Blast furnace molten iron silicon content prediction method and system based on dynamic relation perception

Abstract

The invention relates to the technical field of blast furnaces, and discloses a blast furnace molten iron silicon content prediction method and system based on dynamic relation perception, and the method comprises the steps: obtaining observation data of a blast furnace molten iron silicon content prediction task, and representing the observation data as an irregular multivariable time sequence, a blast furnace molten iron silicon content prediction model comprising an intra-block graph neural network, an inter-block graph neural network and a prediction decoder is constructed, and the intra-block graph neural network obtains block-level representation according to irregular multivariable time sequence mining instantaneous correlation and short-range dependence; the inter-block graph neural network captures a cross-time and cross-variable long-range dependency relationship according to block-level representation to obtain global representation, the prediction decoder obtains a prediction value of the blast furnace molten iron silicon content in a target prediction time period according to the global representation, and the blast furnace molten iron silicon content prediction model after training is completed is used for predicting the blast furnace molten iron silicon content. According to the method, complex blast furnace molten iron silicon content data can be processed, real-time interval prediction in a continuous time period is realized, and the prediction precision is high.

Description

Technical Field

[0001] This invention relates to the field of blast furnace technology, and in particular to a method and system for predicting the silicon content of molten iron in blast furnaces based on dynamic relationship perception. Background Technology

[0002] In blast furnace ironmaking, the silicon content of molten iron is a crucial process parameter reflecting the furnace's thermal state and the quality of the molten iron. Real-time prediction of molten iron silicon content is essential for assessing the blast furnace condition, guiding operational adjustments, and optimizing the smelting process. Traditional methods for predicting molten iron silicon content primarily rely on manual sampling and offline testing, which are not only time-consuming and labor-intensive but also lack real-time accuracy. To improve the real-time performance of molten iron silicon content prediction, two main types of methods have emerged: mechanism-driven methods and data-driven methods.

[0003] Mechanism-driven methods typically model the reaction processes inside blast furnaces based on heat balance, material conservation, and multifluid dynamics theories, generating physically interpretable predictions. However, due to the complexity and variability of the reaction mechanisms inside blast furnaces, mechanistic models often rely on idealized assumptions, making it difficult to meet the real-time prediction needs under dynamic production conditions.

[0004] Data-driven methods can uncover potential patterns by learning from historical production data, enabling nonlinear modeling of silicon content. With the expansion of production data scale and the improvement of computing power, classic machine learning models such as Support Vector Regression (SVR), Random Forest Regression (RFR), and Extreme Gradient Boosting (XGBoost) have been introduced into the prediction of silicon content in molten iron. Support Vector Regression can handle nonlinear relationships using kernel functions. Existing techniques include using SVR to build a silicon content prediction model and combining it with chaotic particle swarm optimization to select optimal parameters. This approach shows certain advantages with small samples, but it requires removing all missing data points of process variables, resulting in the discarding of a large amount of observational information in the original irregular time series. Furthermore, this method only supports single-step prediction, and its prediction time depends on the input sample itself, unable to directly provide the silicon content value at a fixed future time. Ensemble learning models such as Random Forest Regression and Extreme Gradient Boosting have also shown good performance in related research due to their strong feature fitting capabilities. However, these methods typically require a fixed-length, regularized vector as input, necessitating linear interpolation when dealing with irregular multivariate time series. The silicon content in blast furnace hot metal changes abruptly during tapping, but remains relatively stable at other times. Interpolation-based methods smooth out these abrupt changes, leading the model to mistakenly believe the silicon content changes slowly, resulting in the loss of valuable information from the original sequence. Therefore, linear interpolation inevitably introduces noise and information distortion, weakening the reliability of the prediction results. Overall, machine learning models demonstrate high performance in feature fitting and prediction accuracy, but they fundamentally rely on static feature mappings and cannot effectively model dynamic dependencies, making it even more difficult to achieve high-precision predictions of silicon content in irregular multivariate time series blast furnace hot metal.

[0005] With the continuous expansion of industrial data scale and the increasing complexity of its features, data-driven methods have gradually introduced deep learning models such as Back Propagation Neural Networks (BPNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Temporal Convolutional Networks (TCN) into the prediction of silicon content in molten iron. Early attempts were mostly based on BPNN, which can capture the nonlinear relationship between input and output, but it is still insufficient to cope with the complex time dependencies in blast furnace operating conditions. To better model time dependencies, recurrent neural network structures, such as LSTM and GRU, have been widely used. These models effectively learn long-range dependency features through gating mechanisms and have achieved high accuracy in silicon content prediction tasks. However, these models usually still make single-step predictions based on the current input samples, making it difficult to directly output the silicon content value at a specific future time. TCN, with its causal convolution and dilated convolution structures, can capture long-sequence dependencies in parallel and has high computational efficiency, making it suitable for processing long-term blast furnace operating condition sequences. However, in handling missing values, TCN still relies on linear interpolation strategies, which may introduce additional noise. Besides these typical deep learning models, existing technologies also employ feature fusion models, using autoencoders to compress and denoise features, thereby improving prediction stability and accuracy. Customized models for silicon content prediction tasks, such as the Deep Stacked Denoising Autoencoder (D-SDAE), enhance robustness and generalization through layer-by-layer pre-training and denoising mechanisms, making them more adaptable to the noise and missing data problems common in blast furnace environments. However, they still struggle to break free from the paradigm of equally spaced inputs and single-step prediction. Overall, while existing deep learning methods can better integrate features and improve prediction accuracy, these models only support single-point prediction and cannot directly output silicon content values ​​at specific future moments. Furthermore, if predicting silicon values ​​at multiple time points within a range is required, a rolling strategy is commonly used, which can easily lead to error accumulation. Their ability to handle irregularly sampled blast furnace molten iron silicon content data and perform range prediction remains insufficient. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for predicting the silicon content of blast furnace molten iron based on dynamic relationship perception. This method and system can process irregular multivariate time series data on silicon content of blast furnace molten iron, realize interval prediction in continuous time periods, meet the real-time prediction requirements under dynamic production conditions, and improve prediction accuracy.

[0007] To address the aforementioned technical problems, this invention provides a method for predicting the silicon content of blast furnace hot metal based on dynamic relationship perception, comprising: Acquire observational data for the task of predicting silicon content in blast furnace hot metal, and represent the observational data as an irregular multivariate time series; A prediction model for silicon content in molten blast furnace is constructed. The prediction model includes an intra-block graph neural network, an inter-block graph neural network, and a prediction decoder. The intra-block graph neural network mines the instantaneous correlation and short-range dependency between variables based on the irregular multivariate time series to obtain a block-level representation. The inter-block graph neural network captures the long-range dependency relationship across time and across variables based on the block-level representation to obtain a global representation. The prediction decoder obtains the predicted value of silicon content in molten blast furnace within the target prediction time period based on the global representation. The silicon content prediction model for blast furnace molten iron is trained, and the trained blast furnace molten iron silicon content prediction model is used to predict the silicon content of blast furnace molten iron.

[0008] Furthermore, the irregular multivariate time series is as follows: , In the formula, Representing irregular multivariate time series, Indicates variables affecting silicon content , , Represents the total number of variables. The variables include A process variable and a variable silicon, Indicates the variable silicon. Indicates the observation time. Representing variables exist The raw scalar observation data at time 10:00. Representing variables The set of observation times, express The number of observation data.

[0009] Furthermore, the block-level internal graph neural network obtains a block-level representation by mining the instantaneous correlations and short-range dependencies between variables based on the irregular multivariate time series, specifically as follows: The irregular multivariate time series is divided into For the nth, equal-length, non-overlapping blocks... Each block, ,variable The set of observation points is represented as , For variables exist The observation code corresponding to each time point; within the block, an intra-block graph is constructed with observation points as nodes and the lines connecting the nodes as edges; Within the same block, for variables exist Observation code of time and variables exist Observation code of time The observation time difference between the two observation points is expressed as The weights of the corresponding edges are calculated as follows: , In the formula, The corresponding edge weights for the two observation points are: , For learnable weight matrices, , For activation function, Indicates a splicing operation; The silicon-enhanced edge weight variables are calculated as follows: , In the formula, For the silicon-enhanced edge weight variables at the two observation points, It follows a Gaussian distribution. Indicates the variable silicon in The observation code at time; The attention mechanism is used to combine the corresponding edge weights and silicon-enhanced edge weight variables to obtain... Message aggregation information; The observation code is updated using the message aggregation information, resulting in the updated observation code as follows: , In the formula, For the updated , The residual coefficients are used, and GeLU is the activation function. By combining the updated observation codes, the state of the variables over the entire time window is weighted and aggregated to obtain a block-level representation.

[0010] Furthermore, the aforementioned The method for calculating message aggregation information is as follows: , In the formula, express message aggregation information, Indicates traversal All directly connected neighbor nodes, Representing variables In the The set of observation points in each block, Let be the dimension of the key vector. , , , , , This is a learnable weight matrix.

[0011] Furthermore, the weighted aggregation of the variable's state over the entire time window, combined with the updated observation code, yields a block-level representation, specifically as follows: Constructing a time code with practical significance for blast furnace ironmaking: , ; In the formula, for Time encoding of a moment for The time encoding of time in the 1st moment Components in each dimension The dimension of the vector features encoded in time. , Indicates the encoding length. , , , These are learnable parameters; Build query-key-value pairs for temporal semantic attention within the block: = , , ; In the formula, the query vector Representing variables In the Reference points for query time in each block. Representing variables In the Average observation time per block , Representing variables In the A set of observation points in each block; For variables In the The key vector of each block, For variables In the A value vector of blocks, , , It is a learnable weight matrix. For the updated variables In the Each block The observation code at time; The computation block level is represented as: , In the formula, For variables Within the block time period Block-level representation, For attention weights, , This indicates transpose.

[0012] Furthermore, the inter-block graph neural network captures long-range dependencies across time and variables based on the block-level representation to obtain a global representation, specifically as follows: Between blocks, a coarse-grained interaction graph is constructed using the block-level representation of each variable within each block as nodes, the connection between adjacent blocks for the same variable as temporal adjacency edges, and the connection between different variables within adjacent blocks as spatial association edges. Using the same method as the block-based graph neural network for updating observation codes, the target is switched from observation nodes within a block to nodes across blocks, resulting in updated observation codes across blocks. Every two adjacent block nodes are merged into a parent block node as child blocks. The time reference point of the parent block is determined by the weighted average of the number of observations in the child blocks. Recursive aggregation is performed using the same method as the block-level representation computation method described above for the block-level neural network, to obtain any... A global representation of a variable.

[0013] Furthermore, the method for calculating the time reference point of the parent block is as follows: , In the formula, Indicates the aggregation hierarchy. Indicates the first Hierarchical variables In the Each block's time reference point Indicates the first Hierarchical variables In the Each block's time reference point Indicates the first Hierarchical variables In the The time reference point for each block; the first The first under the level The first block is used as the parent block, and the second... The first under the level The first block and the first Each block is used as a sub-block. , Representing variables respectively In the The first block, the first The number of valid observation points within each block.

[0014] Furthermore, the prediction decoder obtains the predicted value of the silicon content of blast furnace hot metal within the target prediction time period based on the global representation, specifically as follows: Using the same method as constructing time codes with practical significance in blast furnace ironmaking, a time representation for each predicted moment is generated, denoted as... , To predict the time, , Predict the target time period; The global representation of the variable silicon in the global representation obtained from the block-based graph neural network and the global representation of silicon are combined. The splicing is used as the input to the decoder, and the output of the decoder is used as the predicted value of the silicon content of blast furnace molten iron within the target prediction time period.

[0015] Furthermore, when training the blast furnace molten iron silicon content prediction model, the loss function is: , In the formula, For loss function, For the sample size, For the sample The set of time points for detecting silicon content in medium. for The true value of silicon content in molten iron at any given moment. for Predicted values ​​of silicon content in blast furnace hot metal at any given time. These are the weighting coefficients for the smoothing term. k1 is the process experience threshold.

[0016] The present invention also provides a blast furnace hot metal silicon content prediction system based on dynamic relationship perception, comprising: The data acquisition and processing module is used to acquire the observation data for the blast furnace molten iron silicon content prediction task and represent the observation data as an irregular multivariate time series. A prediction model construction module is used to construct a prediction model for the silicon content of blast furnace molten iron. The prediction model for the silicon content of blast furnace molten iron includes a block-level internal graph neural network, a block-level inter-graph neural network, and a prediction decoder. The block-level internal graph neural network mines the instantaneous correlation and short-range dependency between variables based on the irregular multivariate time series to obtain a block-level representation. The block-level inter-graph neural network captures the long-range dependency relationship across time and across variables based on the block-level representation to obtain a global representation. The prediction decoder obtains the predicted value of the silicon content of blast furnace molten iron within the target prediction time period based on the global representation. The model training and prediction module is used to train the blast furnace molten iron silicon content prediction model and to predict the silicon content of blast furnace molten iron using the trained blast furnace molten iron silicon content prediction model.

[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention represents the observed data as an irregular multivariate time series. Based on this, a hierarchical block graph neural network framework for predicting the silicon content of blast furnace molten iron is designed. By using intra-block graph neural networks and inter-block graph neural networks in parallel to capture the local dynamic features and global dependencies of the silicon content of blast furnace molten iron, interval prediction is performed over continuous time periods. The generated prediction sequence is not only highly accurate, but also the continuous time series output is precisely matched with the real-time monitoring requirements of blast furnace production processes, which can meet the real-time prediction requirements under dynamic production conditions. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a method in a preferred embodiment of the present invention.

[0019] Figure 2 A schematic diagram of the observation data for the task of predicting the silicon content in blast furnace hot metal.

[0020] Figure 3 This is a diagram illustrating the architecture of a preferred embodiment of the blast furnace molten iron silicon content prediction model. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0022] Reference Figure 1 As shown, this invention discloses a method for predicting the silicon content of blast furnace hot metal based on dynamic relationship perception, comprising the following steps: S1: Obtain the observation data for the blast furnace molten iron silicon content prediction task, represent the observation data as an irregular multivariate time series, and construct the prediction function for the blast furnace molten iron silicon content prediction task.

[0023] like Figure 2 The image shown is a schematic diagram of the observation data for the task of predicting the silicon content in blast furnace hot metal. Figure 2 In the diagram, solid dots represent data present at that time point, while blank dots represent missing values. From... Figure 2 It can be seen that the sampling frequencies of different process variables are not consistent. For example, variable 1 is sampled at 10 minutes, variable 2 at 20 minutes, and the sampling frequency of variable silicon is not fixed. In addition, due to factors such as system maintenance or signal transmission failures, even variables sampled at equal intervals may be missing, thus forming an irregular multivariate time series.

[0024] Furthermore, due to the lag in blast furnace operation, the prediction task needs to infer the silicon content over a future period based on current observation data. For example, it needs to predict the silicon value one hour later (i.e., after 180 minutes) based on data from the previous 120 minutes, thus providing decision support for blast furnace operation. Therefore, each data sample can be divided into three time windows: a historical window of 0–120 minutes, containing observation data that can be used for modeling, with missing data denoted as missing data; an unobservable window of 120–180 minutes, corresponding to data that has not yet occurred at the time of prediction; and a prediction window of 180–240 minutes, which is the target interval that the model needs to output.

[0025] The irregular multivariate time series is: (1), In the formula, Representing irregular multivariate time series, Indicates variables affecting silicon content , , Represents the total number of variables. The variables include A process variable and a variable silicon, Indicates the variable silicon. Indicates the observation time. Representing variables exist The raw scalar observation data at time 10:00. Representing variables The set of observation times, express The number of observed data, variables In the observation time set The content contains One observation data point. From observation time Its corresponding numerical value Together, these variables constitute a complex and irregular multivariate time series structure in the actual operation of a blast furnace, as the observation times of each variable are often non-uniformly distributed and may be accompanied by missing data and asynchronous sampling.

[0026] Given an irregular multivariate time series and the target prediction time period The goal of the problem of predicting silicon content in molten iron is to learn a prediction function. To output The predicted silicon content at each time point is given, therefore the prediction function constructed in this invention is: (2), In the formula, For the prediction model to be learned, It is an irregular multivariate time series. For predictive models in Predicted silicon content at time [time]. The target prediction time period.

[0027] Taking variable silicon and process variables such as material rate and pulverized coal injection rate as examples, for the observation period from 10:00 to 12:00 on April 10, 2025, if the silicon content obtained by sampling at 10:30 is... Then there is The material flow rate was recorded at 10-minute intervals. A total of 12 observation points were set up, and the coal injection rate was recorded at 1-minute intervals. There are a total of 120 observation points, and each observation point and its observed value are recorded. The observation data can then be represented as follows: Furthermore, if the predicted query period is one hour later, i.e., from 13:00 to 14:00 on April 10, 2025, the model needs to output the prediction results for that period. .

[0028] S2: To address the challenges of temporal irregularities, dynamic correlations, low-frequency sparsity of silicon variables, and multi-scale dependencies in silicon content prediction, this invention constructs a system based on dynamic relationship awareness and a segmented graph neural network, as follows: Figure 3 The blast furnace molten iron silicon content prediction model shown includes an intra-block graph neural network, an inter-block graph neural network, and a prediction decoder. The intra-block graph neural network mines the instantaneous correlations and short-range dependencies between variables based on the irregular multivariate time series to obtain a block-level representation. The inter-block graph neural network captures long-range dependencies across time and variables based on the block-level representation to obtain a global representation. The prediction decoder obtains the predicted value of blast furnace molten iron silicon content within the target prediction time period based on the global representation.

[0029] This invention abstracts the silicon content observation sequence as a fully connected graph structure, where the node set consists of the observed data of process variables at each time step, and the edge set represents the connections between any nodes. Considering the complex scenario of irregular sampling of multiple variables in a blast furnace, message passing through a fully connected graph faces challenges of computational complexity and feature obfuscation. To overcome these difficulties, this invention designs a hierarchical block mechanism, employing a hierarchical structure composed of intra-block graph neural networks and inter-block graph neural networks. This structure, through a divide-and-conquer strategy, first meticulously mines the instantaneous correlations and short-range dependencies between variables within local time blocks, eliminating high-frequency noise; then, at the global level, it aggregates the feature representations of different time blocks to capture long-term evolution patterns spanning multiple process stages. This design enhances the model's ability to perceive key trends in silicon content through multi-scale feature extraction, laying a solid foundation for achieving accurate predictions.

[0030] S2-1: The core objective of intra-block modeling is to stably capture immediate dependencies between variables within local time windows, mitigating the impact of irregular sampling and missing data, and providing lower-noise, physically consistent block-level representations for cross-block long-range inference. Within each time block, the model focuses on the observations of the current window, constructs a local graph structure, and aggregates information through a message-passing mechanism to form high-quality block-level representations. These representations, as the basic units for cross-block modeling, avoid the computational bloat caused by direct propagation across the entire time domain and lay the foundation for robustly capturing long-range dependencies, thereby improving the ability to grasp low-frequency trends and enhance model interpretability.

[0031] The block-level neural network obtains a block-level representation by mining the instantaneous correlations and short-range dependencies between variables based on the irregular multivariate time series, specifically as follows: S2-1-1: Divide the irregular multivariate time series into... For the nth, equal-length, non-overlapping blocks... Each block, ,variable The set of observation points is represented as , For variables exist The observation code corresponding to each time point; within the block, an intra-block graph is constructed with observation points as nodes and the lines connecting the nodes as edges (denoted as...). ).

[0032] S2-1-2: To adapt to the complex dynamic coupling characteristics of the blast furnace smelting process, this invention constructs a dynamic relation generator to achieve adaptive adjustment of the relation weights of edges. Within the same block, for variables... exist Observation code of time and variables exist Observation code of time The observation time difference between the two observation points is expressed as The weights of the corresponding edges are calculated as follows: (3), In the formula, The corresponding edge weights for the two observation points are: , For learnable weight matrices, , For activation function, This indicates a splicing operation.

[0033] S2-1-3: Considering the low-frequency characteristics of silicon, a silicon-enhanced edge weight variable is introduced, and the silicon-enhanced edge weight variable is calculated as follows: (4), In the formula, For the silicon-enhanced edge weight variables at the two observation points, As an activation function, it suppresses other irrelevant edges while increasing the message weight of silicon-rich nodes. The linear mapping matrix is ​​a Gaussian distribution, initialized with a zero-mean Gaussian distribution to ensure the original distribution is maintained in the early stages of training, and serves as the input concatenation vector. Indicates the variable silicon in The observation code for the time.

[0034] S2-1-4: Will As the target node, the corresponding edge weights and silicon-enhanced edge weights are obtained through an attention mechanism. The message aggregation information.

[0035] Attention mechanisms are used to aggregate and form target nodes. The message aggregation information includes the comprehensive impact of all neighboring nodes and edge weights on the target node, the target node The method for calculating message aggregation information is as follows: (5), In the formula, express message aggregation information, Indicates traversal All directly connected neighbor nodes, Representing variables In the The set of observation points in each block, Let be the dimension of the key vector. and For query-key vector pairs, For value vectors, , , , , , This is a learnable weight matrix.

[0036] S2-1-5: Considering that the formation of silicon content in molten iron is the comprehensive result of long-term coupling effects of multiple variables within the blast furnace, in order to enrich the node characteristic information and integrate other process variables for the current node... The impact of using the message aggregation information The observation codes are updated to obtain the updated observation codes as follows: (6), In the formula, For the updated , For residual coefficients, the residual terms are... It retains its own historical state information, message items The spatiotemporal interaction information from other process variables is encoded. GeLU is the activation function. The GeLU activation function enhances important features and suppresses noise interference through its asymmetric properties, thereby improving the ability to capture the trend of silicon content change in molten iron.

[0037] S2-1-6: After the target node state is updated, in order to obtain the final representation of the variable, an attention mechanism based on time semantics is designed to decouple the time information into "query time reference point" and "observation time key value". Based on the importance of the time point itself, rather than the size of its observation value, the state of the variable within the entire time window is weighted and aggregated in combination with the updated observation code to obtain a block-level representation.

[0038] S2-1-6-1: Constructing a time code with practical significance for blast furnace ironmaking: , (7); In the formula, for Time encoding of a moment for The time encoding of time in the 1st moment Components in each dimension The dimension of the vector features encoded in time. The size of the value determines the feature representation capability of the encoding. , Indicates the encoding length; this encoding function enables the model to simultaneously model both the trend components and periodic fluctuations in silicon content variation: when A single-layer linear network is used to model the long-term trend of silicon content, ensuring the ability to represent monotonic changes; when ,generate Sine waves of different frequencies were used to fit the periodic patterns of the complex blast furnace ironmaking process. , , , The parameters are learnable, where the frequency parameter is... reciprocal Indicates the actual period length of each periodic component, bit parameter This controls the phase shift of each periodic component, allowing the model to flexibly adjust the starting position of the periodic waveform.

[0039] S2-1-6-2: Constructing query-key-value pairs for temporal semantic attention within a block:

[0040] In the formula, the query vector Representing variables In the Reference points for query time in each block. Representing variables In the Average observation time per block , Representing variables In the A set of observation points in each block; For variables In the The key vector of each block, Indicate the specific observation time; For variables In the A value vector of blocks, Encode the numerical information of the observation nodes at the corresponding time points; , , It is a learnable weight matrix. For the updated variables In the Each block The observation code of time, i.e. , Calculation method and same.

[0041] S2-1-6-3: Computational block-level representation is as follows: (9), In the formula, For variables Within the block time period Block-level representation, For attention weights, Measuring specific moments With block global time reference point Relevance over time The calculation process is separated from the observed values, thereby avoiding the interference of missing values ​​on the weight distribution. The calculation method is as follows: (10), In the formula, This indicates transpose.

[0042] Furthermore, when an observation point is a missing time point, a masking mechanism is used to reset the attention weights corresponding to the missing time point to zero, effectively masking and processing the missing values. This allows the model to automatically identify and increase the importance weights of key process time points, ultimately generating a block-level representation that represents the overall state of the variables throughout the entire time period. .

[0043] S2-2: The core objective of inter-block modeling is to utilize the block-level representation already obtained within the block. Building upon this foundation, the model further captures long-range dependencies across time and variables, and constructs multi-scale temporal abstractions through a hierarchical aggregation strategy, thus obtaining the final representation of each variable within the entire time window. In this way, the model can characterize the slow evolution of silicon content over a longer time period, and effectively control the graph structure size through hierarchical contraction, avoiding the computational burden caused by an overly dense global graph. The final multi-scale representation will serve as the core input for global inference and prediction of molten iron silicon content, providing structured support for capturing the key low-frequency variable, silicon trends.

[0044] The inter-block graph neural network captures long-range dependencies across time and variables based on the block-level representation to obtain a global representation, specifically: S2-2-1: Between blocks, the block-level representation of each variable within each block is used (i.e., , ) is a node, a connection between adjacent blocks for the same variable (i.e. and The lines connecting them are time adjacency edges, and the lines connecting different variables within adjacent blocks (i.e., The lines connecting them ) as spatially related edges to construct a coarse-grained interaction graph (denoted as ) Temporal adjacency edges characterize temporal continuity, while spatially related edges represent their lateral coupling relationship within the same process stage. Through this spatiotemporal coupling edge design, the model can simultaneously perceive vertical dynamic evolution and lateral process collaboration, enabling local information to be fully integrated across the block scale.

[0045] S2-2-2: Using the same method as the block-based graph neural network for updating the observation code (i.e., formulas (3)-(6)), the target is switched from the observation node within the block to the node across the block, and the updated observation code across the block is obtained.

[0046] S2-2-3: After obtaining the updated observation codes across blocks, they need to be further refined into unified global features. To improve the model's temporal abstraction capability, this invention introduces a hierarchical binary aggregation strategy, that is, initializing the features of each independent sub-block and then gradually merging them into globally unique features through recursive merging. Specifically, every two adjacent block nodes are merged into a parent block node as sub-blocks. The time reference point of the parent block is determined by the weighted average of the number of observations within the sub-block. The calculation method for the time reference point of the parent block is as follows: (11), In the formula, Indicates the aggregation hierarchy. Indicates the first Hierarchical variables In the Each block's time reference point Indicates the first Hierarchical variables In the Each block's time reference point Indicates the first Hierarchical variables In the The time reference point for each block; the first The first under the level The first block is used as the parent block, and the second... The first under the level The first block and the first Each block is used as a sub-block; , Representing variables respectively In the The first block, the first The number of valid observation points within each block ensures that the more densely sampled sub-blocks have a greater weight in determining the time reference point, thus naturally possessing the ability to detect missing data; if the number of sub-blocks in the current layer is odd, then the last sub-block is copied to keep the number of sub-blocks even.

[0047] S2-2-4: Next, recursive aggregation is performed using the same method as the block-level representation computation method (i.e., formulas (7)-(9)) used in the block-level neural network. Each layer will... The scale of sub-block features is condensed into Scale of parent block features, this process spans The first scale ultimately yields any second scale. 10 variables (i.e.) The global representation of each process variable and variable silicon (denoted as ). . It integrates information from adjacent time blocks and other variables, thereby achieving a unified representation of local data and global interaction.

[0048] S2-3: To address the low-frequency sparsity of variable silicon, the predictive decoder aims to transform the variable silicon representation obtained from the aforementioned hierarchical graph neural network into a representation that can be predicted at a specified prediction time. The variable silicon prediction value. The prediction decoder obtains the predicted value of silicon content in blast furnace hot metal within the target prediction time period based on the global representation, specifically: S2-3-1: Using the same method as constructing time codes with practical significance for blast furnace ironmaking (i.e., formula (7)), generate the time representation for each predicted moment, denoted as... Used for explicit modeling and prediction of the temporal semantics of time moments. To predict the time, , The target prediction time period.

[0049] S2-3-2: The global representation of the variable silicon in the global representation obtained from the block-based graph neural network (i.e. time )and The concatenation is used as the input to the decoder, and the output of the decoder is used as the predicted value of the silicon content of molten iron in the target prediction time period, that is: (12), In the formula, MLP represents a fully connected structure containing two ReLU activation functions to enhance nonlinear expressive power and capture the complex nonlinear relationship between blast furnace state variables and silicon content; the output layer is linearly activated, making the predicted value... It can freely cover the full range of silicon content fluctuations, reflecting both its slow evolution trend and responding to short-term disturbances.

[0050] S3: Train the blast furnace molten iron silicon content prediction model, and use the trained blast furnace molten iron silicon content prediction model to predict the silicon content of blast furnace molten iron.

[0051] During blast furnace ironmaking, the silicon content in molten iron typically changes slowly and continuously, with very few drastic fluctuations in the short term. Therefore, the predicted results should not exhibit non-physical jumps between adjacent time steps. If only the mean squared error is used as the loss function, although the model may fit the global trend well, it is prone to producing abrupt changes in local time series that do not conform to the process rules, reducing the interpretability and usability of the predicted results.

[0052] To address this issue, and drawing on domain expert knowledge, this invention addresses the upper limit of the rate of change in silicon content under normal production conditions by explicitly introducing a temporal smoothing constraint to prevent unrealistic abrupt changes in the prediction results. Furthermore, a silicon content deviation within 0.05 is considered excellent. Therefore, this invention designs a hierarchical temporal smoothing loss function for predicting silicon content in molten iron and introduces a masking mechanism to handle missing observations, ensuring that loss calculations rely only on valid data points.

[0053] When training the blast furnace molten iron silicon content prediction model, the loss function is: (13), In the formula, For loss function, For the sample size, For the sample The set of time points for detecting silicon content in medium. for The true value of silicon content in molten iron at any given moment. for Predicted values ​​of silicon content in blast furnace hot metal at any given time. These are the weighting coefficients for the smoothing term. k1 is the process experience threshold, and in this embodiment, k1 is set to 0.01.

[0054] The first term in the loss function The standard mean squared error term ensures the absolute accuracy of the prediction results and is used to ensure the predicted values. Compared with the true value The global consistency constraint is the basis of the regression task, ensuring that the model approximates the true observations in the overall trend.

[0055] The second term in the loss function Temporal smoothness is enhanced by penalizing abrupt changes in predicted values ​​at adjacent time steps. Exceed hour, The function is activated by imposing a linear growth penalty, thereby suppressing drastic non-physical fluctuations and making the predicted curve more consistent with the actual blast furnace dynamics.

[0056] This multi-objective loss design, which combines mean square error, physical constraints, and smoothing control, can improve the reliability of predictions within key process thresholds while ensuring prediction stability, providing a more robust decision-making basis for blast furnace process optimization.

[0057] This invention also discloses a predictive system for silicon content in blast furnace hot metal based on dynamic relationship perception, comprising: The data acquisition and processing module is used to acquire the observation data for the blast furnace molten iron silicon content prediction task and represent the observation data as an irregular multivariate time series. A prediction model construction module is used to construct a prediction model for the silicon content of blast furnace molten iron. The prediction model for the silicon content of blast furnace molten iron includes a block-level internal graph neural network, a block-level inter-graph neural network, and a prediction decoder. The block-level internal graph neural network mines the instantaneous correlation and short-range dependency between variables based on the irregular multivariate time series to obtain a block-level representation. The block-level inter-graph neural network captures the long-range dependency relationship across time and across variables based on the block-level representation to obtain a global representation. The prediction decoder obtains the predicted value of the silicon content of blast furnace molten iron within the target prediction time period based on the global representation. The model training and prediction module is used to train the blast furnace molten iron silicon content prediction model and to predict the silicon content of blast furnace molten iron using the trained blast furnace molten iron silicon content prediction model.

[0058] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting the silicon content of molten iron in a blast furnace based on dynamic relationship awareness.

[0059] The present invention also discloses an apparatus including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for predicting the silicon content of blast furnace hot metal based on dynamic relationship awareness.

[0060] The technical challenge of predicting silicon content in blast furnace hot metal lies in effectively extracting and fusing multivariate, multi-scale spatiotemporal features from the irregularly timed original observation sequences and mapping them to the final predicted silicon content value. Specifically: First, the input data exhibits temporal irregularity, with inconsistent sampling time points for different variables and a large number of missing values; second, changes in silicon content are simultaneously influenced by multi-scale factors, including short-term operational fluctuations and long-term furnace condition trends; third, the correlation between process variables and silicon content dynamically changes as the smelting process progresses; and finally, the sparse observation of silicon content itself, the target of prediction, poses a challenge to model learning. These characteristics collectively constitute the core features of this complex time-series prediction problem of blast furnace hot metal silicon content, and dictate that solving this technical challenge requires a modeling method capable of uniformly handling temporal irregularities, multi-scale dependencies, and dynamic correlations.

[0061] To address the technical challenges of predicting silicon content in blast furnace hot metal, this invention represents observed data as an irregular multivariate time series. Based on this, a hierarchical block graph neural network framework for predicting silicon content in blast furnace hot metal is designed. By using intra-block graph neural networks and inter-block graph neural networks in parallel to capture the local dynamic features and global dependencies of silicon content in blast furnace hot metal, interval prediction is performed over continuous time periods. The generated prediction sequence is not only highly accurate, but its continuous time series output also precisely matches the real-time monitoring requirements of blast furnace production processes, thus meeting the real-time prediction needs under dynamic production conditions.

[0062] Compared with the prior art, the advantages of the present invention are as follows: 1. Direct modeling of multivariate native non-uniform sampling data in blast furnace molten iron silicon content prediction was achieved. A hierarchical block graph neural network framework for blast furnace molten iron silicon content prediction was designed, solving the modeling challenge of multi-scale asynchronous time-series data. By directly utilizing real sampling point data, employing a masking mechanism for variable validity labeling, and combining dynamic graph structure construction, incomplete information in the original data was transformed into computable graph structure relationships.

[0063] 2. This invention combines prior knowledge of time decay patterns with graph neural networks. Through time-aware modeling generated by dynamic relationships and the collaborative design of hierarchical block graph neural networks, it achieves hierarchical spatiotemporal feature fusion and realizes end-to-end multi-step prediction. It improves reasoning ability under conditions of incomplete information and exhibits stable predictive performance for multi-scale sampling frequency differences, variable-level data gaps, and system-level time step gaps.

[0064] 3. Breaking through the limitations of traditional point-in-time prediction of molten iron silicon content, it achieves silicon prediction capabilities for extended periods. It expands traditional point-in-time silicon prediction to support continuous time-period silicon prediction, overcoming the limitation that single-point prediction cannot reflect the dynamic trends of process changes.

[0065] 4. Employing a time-aware dynamic decay mechanism, the model can adaptively adjust the contribution weight of historical information; a hierarchical block structure captures the local dynamic features and global dependencies of the blast furnace system in parallel. Unlike traditional methods that only predict single-point silicon values ​​(e.g., only predicting the 180th minute), this invention can generate a continuous one-hour silicon content sequence prediction data within the prediction window, thus intuitively depicting the dynamic trend of future silicon content changes. The generated prediction sequence is not only highly accurate, but its continuous time-series output also precisely matches the real-time monitoring requirements of the blast furnace production process, meeting the real-time prediction needs under dynamic production conditions and providing more comprehensive time-series evolution information for parameter optimization decisions.

[0066] 5. To address the continuous, gradual variation and low frequency of silicon content in blast furnace hot metal, an optimization mechanism combining a time-series smoothing loss function and gating enhancement was designed. Based on the continuous and gradual nature of silicon content changes in blast furnace hot metal, a time-series smoothing loss function was designed to improve prediction accuracy. This function ensures that the prediction results conform to the physical gradual variation law of silicon content while effectively maintaining the model's prediction accuracy. Simultaneously, considering the importance and low frequency of silicon content as a key quality indicator, the model significantly increased the weight allocation of silicon content features in the gating mechanism, enabling each layer of the network to effectively focus on the feature patterns with the greatest process guidance value. This dual optimization strategy, integrating industrial laws and key features, ensures that the output results meet actual process requirements while effectively characterizing complex nonlinear relationships, achieving a complementary advantage between process knowledge and data-driven methods.

[0067] To further verify the advantages of the present invention, this embodiment uses data collected on a blast furnace to conduct experiments, including a series of comparative and ablation experiments. The hardware configuration includes a 6-core Intel Xeon 5118 processor, 32GB of RAM, and an NVIDIA GeForce GTX 2080Ti graphics card; the software environment is based on CUDA 11.8 and Python 3.10, built using the PyTorch 2.0.0 deep learning framework.

[0068] The performance evaluation metrics used are Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Hit Rate (HR) for silicon content prediction, which measure the difference between the model's predicted values ​​and the actual test values. The smaller the RMSE and MAE, the higher the model's prediction accuracy and the better its performance; the larger the HR, the higher the hit rate and the stronger the model's predictive ability. Based on the experience of on-site experts, an absolute error between the predicted and actual test values ​​within 0.1% is considered acceptable.

[0069] After removing abnormal data caused by blast furnace shutdowns and equipment failures from the blast furnace historical database, a basic dataset with temporal integrity was formed, consisting of 185,761 data points. The dataset was divided into training, validation, and test sets in a ratio of 7.5:1.5:1. To verify the advantages of the method of this invention, it was compared with eight existing models for predicting silicon content in molten iron, including classic machine learning models SVR, XGBoost, and RFR, and deep learning models BPNN, LSTM, GRU, and TCN. All baseline models maintained the same parameter configuration. In addition, the D-SDAE model, specifically constructed for predicting silicon content in molten iron, was also compared. The experimental results are shown in Table 1.

[0070] Table 1. Performance Comparison of Different Models in Predicting Silicon Content in Molten Iron

[0071] As shown in Table 1, the performance of the method of this invention is superior to other methods in all indicators. Particularly noteworthy is that while the D-SDAE model, which has the best HR performance among existing methods, achieves a hit rate of 90.13%, its RMSE (0.06142) and MAE (0.05075) are still significantly higher than those of the method of this invention. This indicates that traditional methods have inherent limitations in handling cross-scale dependencies in the prediction of molten iron silicon content. These limitations mainly stem from two aspects: firstly, conventional time-series prediction models struggle to effectively capture the irregular and nonlinear characteristics of molten iron silicon content changes; secondly, while existing feature fusion methods attempt to integrate multi-scale information, they often process irregular data through interpolation operations, which may distort the original data distribution and introduce additional errors.

[0072] The hierarchical block graph neural network in this invention effectively solves the cross-scale dependency modeling problem in predicting silicon content in molten iron through an innovative layer collaboration mechanism within and between blocks. Its data processing strategy, which only processes data from actual observation points, avoids the error accumulation problem introduced by traditional interpolation methods. Furthermore, addressing the unique nonlinear dynamic characteristics of silicon content in molten iron, this invention accurately captures key feature change patterns by reconstructing the loss function structure. While achieving the more challenging task of multi-step prediction over 60-120 minutes, it maintains a high prediction hit rate and keeps RMSE and MAE at excellent levels. In addition, the method's prediction errors are concentrated near zero with few extreme errors, further verifying its stability and reliability in complex industrial scenarios. It provides prediction results for blast furnace smelting processes that combine numerical accuracy and trend interpretability, significantly improving the predictability and scientific rigor of furnace condition control.

[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception, characterized in that, include: Acquire observational data for the task of predicting silicon content in blast furnace hot metal, and represent the observational data as an irregular multivariate time series; A prediction model for silicon content in molten blast furnace is constructed. The prediction model includes an intra-block graph neural network, an inter-block graph neural network, and a prediction decoder. The intra-block graph neural network mines the instantaneous correlation and short-range dependency between variables based on the irregular multivariate time series to obtain a block-level representation. The inter-block graph neural network captures the long-range dependency relationship across time and across variables based on the block-level representation to obtain a global representation. The prediction decoder obtains the predicted value of silicon content in molten blast furnace within the target prediction time period based on the global representation. The silicon content prediction model for blast furnace molten iron is trained, and the trained blast furnace molten iron silicon content prediction model is used to predict the silicon content of blast furnace molten iron.

2. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to claim 1, characterized in that: The irregular multivariate time series is: ， In the formula, Representing irregular multivariate time series, Indicates variables affecting silicon content , , Represents the total number of variables. The variables include A process variable and a variable silicon, Indicates the variable silicon. Indicates the observation time. Representing variables exist The raw scalar observation data at time 10:

00. Representing variables The set of observation times, express The number of observation data.

3. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to claim 2, characterized in that: The block-level neural network obtains a block-level representation by mining the instantaneous correlations and short-range dependencies between variables based on the irregular multivariate time series, specifically as follows: The irregular multivariate time series is divided into For the nth, equal-length, non-overlapping blocks... Each block, ,variable The set of observation points is represented as , For variables exist The observation code corresponding to each time point; within the block, an intra-block graph is constructed with observation points as nodes and the lines connecting the nodes as edges; Within the same block, for variables exist Observation code of time and variables exist Observation code of time The observation time difference between the two observation points is expressed as The weights of the corresponding edges are calculated as follows: ， In the formula, The corresponding edge weights for the two observation points are: , For learnable weight matrices, , For activation function, Indicates a splicing operation; The silicon-enhanced edge weight variables are calculated as follows: ， In the formula, For the silicon-enhanced edge weight variables at the two observation points, It follows a Gaussian distribution. Indicates the variable silicon in The observation code at time; The attention mechanism is used to combine the corresponding edge weights and silicon-enhanced edge weight variables to obtain... Message aggregation information; The observation code is updated using the message aggregation information, resulting in the updated observation code as follows: ， In the formula, For the updated , The residual coefficients are used, and GeLU is the activation function. By combining the updated observation codes, the state of the variables over the entire time window is weighted and aggregated to obtain a block-level representation.

4. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to claim 3, characterized in that: The The method for calculating message aggregation information is as follows: ， In the formula, express message aggregation information, Indicates traversal All directly connected neighbor nodes, Representing variables In the The set of observation points in each block, Let be the dimension of the key vector. , , , , , This is a learnable weight matrix.

5. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to claim 3, characterized in that: The block-level representation is obtained by weighting and aggregating the states of variables throughout the entire time window using the updated observation codes. Constructing a time code with practical significance for blast furnace ironmaking: ， ； In the formula, for Time encoding of a moment for The time encoding of time in the 1st moment Components in each dimension The dimension of the vector features encoded in time. , Indicates the encoding length. , , , These are learnable parameters; Build query-key-value pairs for temporal semantic attention within the block: = ， ， ； In the formula, the query vector Representing variables In the Reference points for query time in each block. Representing variables In the Average observation time per block , Representing variables In the A set of observation points in each block; For variables In the The key vector of each block, For variables In the A value vector of blocks, , , It is a learnable weight matrix. For the updated variables In the Each block The observation code at time; The computation block level is represented as: ， In the formula, For variables Within the block time period Block-level representation, For attention weights, , This indicates transpose.

6. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to claim 5, characterized in that: The inter-block graph neural network captures long-range dependencies across time and variables based on the block-level representation to obtain a global representation, specifically: Between blocks, a coarse-grained interaction graph is constructed using the block-level representation of each variable within each block as nodes, the connection between adjacent blocks for the same variable as temporal adjacency edges, and the connection between different variables within adjacent blocks as spatial association edges. Using the same method as the block-based graph neural network for updating observation codes, the target is switched from observation nodes within a block to nodes across blocks, resulting in updated observation codes across blocks. Every two adjacent block nodes are merged into a parent block node as child blocks. The time reference point of the parent block is determined by the weighted average of the number of observations in the child blocks. Recursive aggregation is performed using the same method as the block-level representation computation method described above for the block-level neural network, to obtain any... A global representation of a variable.

7. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to claim 6, characterized in that: The method for calculating the time reference point of the parent block is as follows: ， In the formula, Indicates the aggregation hierarchy. Indicates the first Hierarchical variables In the Each block's time reference point Indicates the first Hierarchical variables In the Each block's time reference point Indicates the first Hierarchical variables In the The time reference point for each block; the first The first under the level The first block is used as the parent block, and the second... The first under the level The first block and the first Each block is used as a sub-block. , Representing variables respectively In the The first block, the first The number of valid observation points within each block.

8. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to claim 5, characterized in that: The prediction decoder obtains the predicted value of silicon content in blast furnace hot metal within the target prediction time period based on the global representation, specifically as follows: Using the same method as constructing time codes with practical significance in blast furnace ironmaking, a time representation for each predicted moment is generated, denoted as... , To predict the time, , Predict the target time period; The global representation of the variable silicon in the global representation obtained from the block-based graph neural network and the global representation of silicon are combined. The splicing is used as the input to the decoder, and the output of the decoder is used as the predicted value of the silicon content of blast furnace molten iron within the target prediction time period.

9. The method for predicting silicon content in blast furnace hot metal based on dynamic relationship perception according to any one of claims 1-8, characterized in that: When training the blast furnace molten iron silicon content prediction model, the loss function is: ， In the formula, For loss function, For the sample size, For the sample The set of time points for detecting silicon content in medium. for The true value of silicon content in molten iron at any given moment. for Predicted values ​​of silicon content in blast furnace hot metal at any given time. These are the weighting coefficients for the smoothing term. k1 is the process experience threshold.

10. A blast furnace hot metal silicon content prediction system based on dynamic relationship perception, characterized in that, include: The data acquisition and processing module is used to acquire the observation data for the blast furnace molten iron silicon content prediction task and represent the observation data as an irregular multivariate time series. A prediction model construction module is used to construct a prediction model for the silicon content of blast furnace molten iron. The prediction model for the silicon content of blast furnace molten iron includes a block-level internal graph neural network, a block-level inter-graph neural network, and a prediction decoder. The block-level internal graph neural network mines the instantaneous correlation and short-range dependency between variables based on the irregular multivariate time series to obtain a block-level representation. The block-level inter-graph neural network captures the long-range dependency relationship across time and across variables based on the block-level representation to obtain a global representation. The prediction decoder obtains the predicted value of the silicon content of blast furnace molten iron within the target prediction time period based on the global representation. The model training and prediction module is used to train the blast furnace molten iron silicon content prediction model and to predict the silicon content of blast furnace molten iron using the trained blast furnace molten iron silicon content prediction model.

Images