A method and system for predicting load and power in the steel industry based on a spatiotemporal graph fusion model.

By constructing a spatiotemporal graph fusion model with Transformer temporal channel and heterogeneous graph convolutional spatial channel, the problems of insufficient equipment topology association modeling and limited long-term dependent feature extraction capability in load forecasting of the steel industry are solved, achieving high-precision load forecasting, which is suitable for flexible power consumption and low-carbon optimized scheduling in the steel industry.

CN122697311BActive Publication Date: 2026-11-10SOUTHEAST UNIV +2
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611184109.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-06
Publication Date
2026-11-10
Estimated Expiration
2046-08-06

AI Technical Summary

Technical Problem

Existing load forecasting methods for the steel industry lack consideration for production process constraints and equipment coordination relationships, making it difficult to reflect the real load change patterns under complex industrial production scenarios. Furthermore, they are not robust to abnormal operating conditions and have insufficient forecasting accuracy.

Method used

A spatiotemporal graph fusion model is adopted to construct a Transformer temporal channel and a heterogeneous graph convolutional spatial channel. By constructing a task-device heterogeneous graph, temporal and topological features are extracted, and feature fusion is performed through a self-attention mechanism to output future load prediction results.

Benefits of technology

It improves the accuracy and robustness of load forecasting in the steel industry, adapts to complex production scenarios, and supports flexible power consumption, demand response, and low-carbon optimized scheduling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122697311B_ABST
    Figure CN122697311B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for predicting load power in the steel industry based on a spatiotemporal graph fusion model. The method includes: acquiring the operating status and historical power sequences of key equipment in the steel industry to construct a production task-equipment heterogeneous graph; extracting long-term temporal dependency features of the load sequence through a Transformer time channel to capture intraday load fluctuations and cross-cycle trends; characterizing the topological relationships between tasks and equipment through a heterogeneous graph convolutional spatial channel, and performing message passing and feature aggregation on power interactions between equipment; fusing the outputs of the time and spatial channels, and performing weighted aggregation through a self-attention mechanism to form the final high-dimensional fused features; inputting the fused features into a fully connected network to output the future power sequence, thus achieving power prediction for the steel industry. This invention improves the accuracy of power prediction and robustness to abnormal operating conditions, supporting the realization of flexible scheduling and demand response.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of power load forecasting and industrial intelligent scheduling technology, and in particular relates to a method for predicting short-process production power in the steel industry that integrates Transformer temporal modeling and heterogeneous graph convolutional topology modeling. Background Technology

[0002] With the continuous development of new power systems, high-energy-consuming industrial users are gradually becoming important participants in the flexible regulation of the power system. Among them, the steel industry, as a typical process manufacturing industry, is characterized by continuous production processes, high equipment power, close process coupling, and significant load fluctuations. Its electricity consumption behavior not only directly affects the production costs of enterprises, but also relates to the peak-valley regulation capacity of the power grid and the effectiveness of demand response.

[0003] In the short-process production of the steel industry, key equipment such as EAF (Electric Arc Furnace), AOD (Argon Oxygen Decarburization Furnace), LF (Ladle Furnace), and CC (Continuous Casting Machine) are subject to strict process sequence constraints and resource coupling relationships. The coordinated operation of different equipment leads to power curves exhibiting significant nonlinearity, time-varying characteristics, and complex fluctuations. Therefore, accurately predicting the load power of the steel industry has become an important issue in industrial load optimization scheduling and flexible power consumption research.

[0004] Currently, research on industrial load forecasting mainly employs deep learning methods such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Networks (CNN) to mine time-series patterns from historical load data. While these methods can improve forecast accuracy to some extent, the following problems still exist:

[0005] 1. Most existing methods are based solely on historical power sequences for modeling, lacking consideration of steel production process constraints and equipment coordination relationships, and thus failing to reflect the real load variation patterns under complex industrial production scenarios;

[0006] 2. Traditional time series models are difficult to simultaneously characterize the long-term dependence of load in the steel industry and the spatial topological relationship between equipment, and are not adaptable to power fluctuations caused by changes in production processes.

[0007] 3. Existing models are not robust to abnormal operating conditions such as equipment shutdown, temporary furnace addition, and process switching, and are prone to increased prediction errors in complex industrial scenarios.

[0008] 4. Existing methods lack modeling of the structured relationship between "task-equipment-process flow" and cannot effectively integrate production mechanism constraints with data-driven feature extraction capabilities.

[0009] In recent years, GNNs (Graph Neural Networks) have demonstrated good performance in modeling complex topological relationships and can be used to describe the relationships between industrial equipment. Transformer models, on the other hand, have significant advantages in extracting long-term temporal features and can effectively capture long-term dependency features in load sequences. However, most existing studies only use a single model for load forecasting and have not yet formed a unified forecasting framework that can simultaneously integrate temporal features, topological features, and industrial process constraints.

[0010] Therefore, it is urgent to propose a load power prediction model and system for the steel industry based on a spatiotemporal graph fusion model to solve the problems of insufficient spatiotemporal feature coupling ability, insufficient industrial topology modeling, and weak adaptability to abnormal operating conditions in the existing technology. This will improve the power prediction accuracy and robustness in complex production scenarios of the steel industry and provide technical support for flexible scheduling, demand response, and low-carbon operation of the steel industry. Summary of the Invention

[0011] The technical problem to be solved by this invention is to provide a method and system for predicting load power in the steel industry based on a spatiotemporal graph fusion model, which addresses the shortcomings of existing methods for predicting load in the steel industry, such as insufficient modeling of equipment topology associations, limited ability to extract long-term time-series dependent features, and weak prediction robustness under complex operating conditions. This method can be applied to scenarios such as flexible power consumption, demand response, industrial load prediction, and low-carbon optimization scheduling in the steel industry.

[0012] To solve the above technical problems, the present invention is achieved through the following technical solution:

[0013] First, this invention proposes a method for predicting the load power of the steel industry based on a spatiotemporal graph fusion model, comprising the following steps:

[0014] S1 acquires multi-dimensional industrial data from short-process production in the steel industry, including equipment operating status data, historical power sequence data, production task execution data, time-of-use electricity price data, and timestamp data.

[0015] S2, construct a heterogeneous graph of steel production tasks and equipment, using nodes to represent production tasks and production equipment, and using edges to represent equipment occupancy relationships, process flow relationships and temporal dependencies;

[0016] S3, Construct a time channel based on a Transformer encoder to extract time-series features from historical power sequences in order to obtain long-term time-series dependency features in the load sequence;

[0017] S4 constructs a spatial channel based on a heterogeneous graph convolutional network to model the topological relationship between production tasks and equipment, and realizes feature propagation between nodes through a message passing mechanism;

[0018] S5 fuses the output features of the time channel and the output features of the spatial channel, and performs weighted aggregation through a self-attention mechanism to form fused features;

[0019] S6 will integrate the feature input power prediction layer and output the power prediction results of the steel industry load for future time periods.

[0020] Meanwhile, this invention also proposes a load power prediction system for the steel industry based on a spatiotemporal graph fusion model, comprising:

[0021] The data acquisition module is used to acquire equipment operating status data, historical power sequence data, production task data, and electricity price data in the short-process production of the steel industry.

[0022] The heterogeneity graph construction module is used to construct task-device heterogeneity graphs;

[0023] The temporal feature extraction module is used to extract temporal features based on the Transformer encoder;

[0024] The spatial feature extraction module is used to extract device topological association features based on heterogeneous graph convolutional networks;

[0025] The spatiotemporal fusion module is used to fuse temporal and spatial features;

[0026] The power prediction module is used to output power prediction results for the steel industry in the future time period.

[0027] Finally, the present invention proposes a computer storage medium storing a computer program that, when run by a processor, can execute the above-described method.

[0028] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0029] 1. The load power prediction method for the steel industry based on the spatiotemporal graph fusion model in this invention realizes the joint modeling of the load time characteristics and equipment topology characteristics of the steel industry by constructing a Transformer time channel and a heterogeneous graph convolutional spatial channel, which can effectively improve the power prediction accuracy in complex industrial scenarios.

[0030] 2. This invention improves the model's adaptability to industrial production processes and its physical consistency by constructing a task-equipment heterogeneous graph, which incorporates the constraints of steel industry production processes, equipment coordination relationships, and task timing relationships into the power prediction model.

[0031] 3. The method of the present invention is applicable to scenarios such as flexible power use, demand response, low-carbon optimized scheduling, and industrial load forecasting in the steel industry, and can provide technical support for cost reduction and carbon reduction in the steel industry and safe and economical operation of the power system. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the steel industry load power prediction method of the present invention.

[0034] Figure 2 This is a schematic diagram of the power prediction model structure of the spatiotemporal graph fusion model.

[0035] Figure 3 This is a diagram of the steel production task and equipment heterogeneity of the present invention.

[0036] Figure 4 It is a graph showing the changes in model training loss and validation loss.

[0037] Figure 5 This is a schematic diagram comparing the actual power and predicted power of a typical day's steel load. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1: As Figure 1 As shown, this invention proposes a method for predicting the load power of the steel industry based on a spatiotemporal graph fusion model, comprising the following steps:

[0040] S1 acquires multi-dimensional industrial data from short-process production in the steel industry, including equipment operating status data, historical power sequence data, production task execution data, time-of-use electricity price data, and timestamp data.

[0041] Among them, short-process production technology in the steel industry, such as Figure 2 As shown, key equipment includes the electric arc furnace (EAF), the argon-oxygen decarburization furnace (AOD), the ladle furnace (LF), and the continuous casting machine (CC).

[0042] S2, Construct a steel production task-equipment heterogeneity diagram;

[0043] like Figure 3 As shown, nodes represent production task nodes and equipment nodes, and edges represent equipment occupancy relationships, process flow relationships, and timing dependencies.

[0044] S3. Construct a time channel based on a Transformer encoder to extract time-series features from historical power sequences.

[0045] S4 constructs a spatial channel based on a heterogeneous graph convolutional network and realizes feature propagation between nodes through a message passing mechanism.

[0046] S5 fuses the output features of the time channel and the output features of the spatial channel, and forms a fused feature through a self-attention mechanism.

[0047] S6 will integrate the feature input power prediction layer and output the power prediction results of the steel industry load for future time periods.

[0048] Furthermore, the Transformer encoder includes a feature embedding layer, a position encoding layer, a multi-head self-attention layer, and a feedforward neural network layer;

[0049] The expression for the feature embedding process is:

[0050] ,

[0051] In the formula, The input is the industrial feature matrix, representing the raw industrial data at time step t; The embedding weight matrix is ​​used to map the original input to a high-dimensional feature space; For the embedding layer bias term; This is the high-dimensional feature matrix after embedding.

[0052] Furthermore, the position encoding adopts a sine-cosine encoding method, the expression of which is:

[0053] ,

[0054] ,

[0055] In the formula, For even-dimensional positional encoding values; is the odd-dimensional position encoding value; pos is the time step position index; i is the feature dimension index; d is the total dimension of the position encoding.

[0056] Furthermore, the expression for the multi-head self-attention mechanism is:

[0057] ,

[0058] In the formula, Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; The dimension of the key vector; This is a normalization function used to calculate attention weights; This is the output of self-attention.

[0059] Furthermore, the heterogeneous graph convolutional network includes task nodes, device nodes, device occupancy edges, process flow edges, and temporal dependency edges;

[0060] in:

[0061] Task nodes are used to represent production tasks in the steel production process;

[0062] Equipment nodes are used to represent electric arc furnace (EAF), argon-oxygen decarburization furnace (AOD), ladle furnace (LF), and continuous casting machine (CC);

[0063] The device occupancy edge is used to describe the occupancy relationship between tasks and devices;

[0064] The process flow edge is used to describe the sequential relationship of steel production processes;

[0065] Temporal dependency edges are used to describe the temporal continuity between production tasks.

[0066] Furthermore, the heterogeneous graph convolutional network updates node features through a message passing mechanism, and its message generation expression is:

[0067] ,

[0068] In the formula, This is the message vector that neighbor node u in layer l transmits to node v. The feature representation of node v in layer l; The feature representation of the neighbor node u in the l-th layer; This is the message passing weight matrix.

[0069] The message aggregation expression is:

[0070] ,

[0071] In the formula, The message representation of node v after aggregation at level l; It is a mean aggregation function; The set of messages passed to node v by all neighboring nodes.

[0072] The node update expression is:

[0073] ,

[0074] In the formula, This represents the feature representation of node v after the update at layer l+1. For activation functions; For node bias terms; These are the characteristics of the aggregated message.

[0075] Furthermore, the temporal channel output features and spatial channel output features are fused using a concatenation and nonlinear mapping method, and its expression is as follows:

[0076] ,

[0077] In the formula, Output features for the time channel; F represents the spatial channel output feature; F represents the spatiotemporal fusion feature.

[0078] The fused features undergo a nonlinear transformation via a fully connected network, the expression of which is:

[0079] ,

[0080] In the formula, For the final fusion output features; , This is the weight matrix of the fully connected layer; , For bias terms; is the linear rectified activation function; F is the fused feature after splicing.

[0081] Furthermore, a self-attention weighted aggregation mechanism is employed to enhance features at key time steps, the expression of which is:

[0082] ,

[0083] In the formula, Let be the attention weight corresponding to the i-th time step.

[0084] The aggregated context vector expression is:

[0085] ,

[0086] In the formula, Let be the feature vector corresponding to the i-th time step.

[0087] Furthermore, the future power prediction result is output through a fully connected network, and its expression is:

[0088] ,

[0089] In the formula, C represents the predicted power sequence; C is the context aggregation vector. This is the weight matrix of the fully connected layer. This is used to output the bias vector of the fully connected layer.

[0090] Furthermore, in step S6, the power prediction layer adopts a multi-layer fully connected network to map the context vector obtained by spatiotemporal feature fusion and self-attention weighted aggregation into a power prediction sequence for the steel industry load at multiple consecutive time steps in the future.

[0091] The specific processing steps of the power prediction layer include:

[0092] S61, the context aggregation vector C is input into the first fully connected layer for dimension mapping, and a nonlinear feature transformation is performed through a linear rectified activation function to obtain the first hidden feature, the expression of which is:

[0093] ,

[0094] In the formula, This is the first hidden feature; This is the weight matrix of the first fully connected layer; This is the bias vector for the first fully connected layer.

[0095] S62, the first hidden feature The second fully connected layer is input to further extract higher-order features related to future power changes, resulting in the second hidden feature, whose expression is:

[0096] ,

[0097] In the formula, This is the second hidden feature; This is the weight matrix of the second fully connected layer; This is the bias vector for the second fully connected layer.

[0098] S63, the second hidden feature The input-output fully connected layer maps the power prediction sequence for a future time period, and its expression is:

[0099] ,

[0100] ,

[0101] In the formula, This is the normalized power prediction sequence output by the power prediction layer; To output the weight matrix of the fully connected layer; To output the bias vector of the fully connected layer; The number of prediction time slots within the time period to be predicted; This represents the normalized power prediction value corresponding to the k-th future time slot. The number of neurons in the output fully connected layer is set to... This allows each output neuron to correspond to a future prediction time slot.

[0102] S64, when the input power data is standardized before model training, the normalized prediction result output by the power prediction layer is destandardized to obtain a prediction result with actual power units. Its expression is:

[0103] ,

[0104] In the formula, This represents the predicted load power of the steel industry in the k-th time slot. and These represent the mean and standard deviation of the power samples in the training data, respectively. If the input power data is not standardized, the output of the power prediction layer is directly used as the power prediction result for the corresponding time period.

[0105] Arrange the denormalized predicted power values ​​for each time slot in chronological order to obtain the predicted power load sequence for the steel industry in the future time period:

[0106] ,

[0107] In this embodiment, the lengths of both the input and output sequences are set to 48, and the time interval for each time slot is 30 minutes. Therefore, based on the multidimensional industrial data of the previous 24 hours, the model outputs the predicted load power of the steel industry for 48 time slots within the next 24 hours. The feature dimension of the context aggregation vector is 256, the first and second hidden layers of the power prediction layer are set to 128 and 64 dimensions respectively, and the output layer dimension is set to 48.

[0108] The weight matrix , , and bias vector , , Instead of being pre-set, the power is obtained through error backpropagation during model training. Specifically, the predicted power sequence output by the power prediction layer is compared with the actual power sequence for the corresponding future time period. The Huber loss function is used to calculate the prediction error, and the AdamW optimizer is used to jointly iteratively update the network parameters of the time channel, spatial channel, spatiotemporal fusion module, and power prediction layer. Once the validation set loss stabilizes or reaches the preset number of training rounds, the trained network parameters are fixed. In the actual prediction process, real-time acquired industrial data is input into the trained model, and the predicted load power of the steel industry for the future time period can be obtained through a single forward propagation.

[0109] Furthermore, the Huber loss function is used to train the model to improve its robustness under abnormal operating conditions such as equipment failure, temporary shutdown, and dynamic adjustment of production tasks.

[0110] Next, the method proposed in Example 1 will be verified by simulation.

[0111] The experiment used actual operating data from a steel company as the test dataset. The experimental environment consisted of a Windows 11 operating system, an Intel Core i7-13700KF CPU, and an NVIDIA RTX3080 GPU. The model was implemented using Python within the Tensorflow framework. The model training parameters are shown in Table 1.

[0112] Table 1 Training parameter settings for the model of this invention

[0113] parameter numerical values Optimizer AdamW Number of training rounds 10 Input / output sequence length 48 (30 min / slot) Batch size 32 Hidden layer dimensions 256 Initial learning rate 0.001 Weight decay <![CDATA[1×10⁻ 5 ]]> Bullish attention head count 8 Dropout rate 0.1 Graph Convolution Type GAT Transformer layer number 3

[0114] First, the training convergence and stability of the model of this invention are verified. The changes in training set loss and validation set loss during model training are recorded, and the results are as follows: Figure 4 As shown, with the increase in the number of training rounds, the training loss gradually decreased from approximately 0.51 to around 0.12, and the validation loss decreased from approximately 0.26 to around 0.10, and then stabilized in the later stages of training. The training loss and validation loss showed consistent trends, without significant divergence or drastic fluctuations, indicating that the model of this invention can effectively learn the temporal changes in load and equipment topology characteristics in the steel industry, and has good convergence and generalization capabilities.

[0115] Secondly, the predictive ability of the model of this invention for intraday load changes was verified. Five typical working days were randomly selected from the test set, and the predicted power of the model was compared with the actual power. The results are as follows: Figure 5 As shown. This invention can accurately track the load change trend in the steel industry and can predict load peak and trough changes relatively well.

[0116] In summary, this invention achieves joint modeling of the spatiotemporal characteristics of power in the steel industry by integrating the Transformer temporal channel and the heterogeneous graph convolutional spatial channel. This can effectively improve the load prediction accuracy in complex industrial scenarios and provide technical support for flexible scheduling, demand response, and low-carbon operation in the steel industry.

[0117] Example 2: Based on the method proposed in Example 1, this example proposes a load power prediction system for the steel industry based on a spatiotemporal graph fusion model, including:

[0118] The data acquisition module is used to acquire multi-dimensional industrial data in the short-process production of the steel industry, including equipment operating status data, historical power sequence data, production task execution data, time-of-use electricity price data, and timestamp data.

[0119] The heterogeneous graph construction module is used to construct a heterogeneous graph of steel production tasks and equipment. Nodes represent production tasks and production equipment, and edges represent equipment occupancy relationships, process flow relationships, and temporal dependencies.

[0120] The time feature extraction module is used to construct a time channel based on the Transformer encoder and extract time-series features from the historical power sequence to obtain long-term time-series dependency features in the load sequence.

[0121] The spatial feature extraction module is used to construct a spatial channel based on a heterogeneous graph convolutional network, model the topological relationship between production tasks and equipment, and realize feature propagation between nodes through a message passing mechanism.

[0122] The spatiotemporal fusion module is used to fuse the output features of the temporal channel and the output features of the spatial channel, and performs weighted aggregation through a self-attention mechanism to form fused features;

[0123] The power prediction module is used to input the fused features into the power prediction layer and output the power prediction results of the steel industry load for a future time period.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for predicting load power in the steel industry based on a spatiotemporal graph fusion model, characterized in that, Includes the following steps: Acquire multi-dimensional industrial data from short-process production in the steel industry, including equipment operating status data, historical power sequence data, production task execution data, time-of-use electricity price data, and timestamp data; Construct a heterogeneous graph of steel production tasks and equipment, where nodes represent production tasks and production equipment, and edges represent equipment occupancy relationships, process flow relationships, and temporal dependencies. A time channel based on a Transformer encoder is constructed to extract time-series features from historical power sequences in order to obtain long-term time-series dependency features in the load sequence. A spatial channel based on heterogeneous graph convolutional networks is constructed to model the topological relationship between production tasks and equipment, and feature propagation between nodes is realized through a message passing mechanism; The heterogeneous graph includes: Task nodes are used to represent production tasks in the steel production process; Equipment nodes are used to represent electric arc furnace (EAF), argon-oxygen decarburization furnace (AOD), ladle furnace (LF), and continuous casting machine (CC). Device occupancy edge is used to describe the occupancy relationship between tasks and devices; The process flow edge is used to describe the sequential relationship of the production process; Temporal dependency edges are used to describe the temporal continuity between production tasks; The output features of the time channel and the output features of the spatial channel are fused and weighted by a self-attention mechanism to form a fused feature; The fusion features are input into the power prediction layer, which outputs the power load prediction results for the steel industry in the future time period.

2. The method according to claim 1, characterized in that, The time channel adopts a Transformer encoder structure, and its input features include power data, time-of-use electricity price data, and timestamp data. The Transformer encoder includes a feature embedding layer, a position encoding layer, a multi-head self-attention layer, and a feedforward neural network layer.

3. The method according to claim 2, characterized in that, The process expression for the feature embedding layer is as follows: , In the formula, The input is the industrial feature matrix, representing the raw industrial data at time step t; The embedding weight matrix is ​​used to map the original input to a high-dimensional feature space; For the embedding layer bias term; This is the high-dimensional feature matrix after embedding; The position coding layer adopts a sine-cosine position coding method, the expression of which is: , , In the formula, For even-dimensional positional encoding values; is the odd-dimensional position encoding value; pos is the time step position index; i is the feature dimension index; d is the total dimension of the position encoding.

4. The method according to claim 3, characterized in that, The expression for the multi-head attention mechanism of the multi-head self-attention layer is: , In the formula, Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; The dimension of the key vector; This is a normalization function used to calculate attention weights; This is the output of self-attention.

5. The method according to claim 4, characterized in that, Heterogeneous graph convolutional networks update node features through a message passing mechanism, and their message generation expression is: , In the formula, This is the message vector that neighbor node u in layer l transmits to node v. The feature representation of node v in layer l; The feature representation of the neighbor node u in the l-th layer; This is the message passing weight matrix; The message aggregation expression is: , In the formula, The message representation of node v after aggregation at level l; It is a mean aggregation function; The set of messages passed from all neighboring nodes to node v; The node feature update expression is: , In the formula, This represents the feature representation of node v after the update at layer l+1. For activation functions; For node bias terms; These are the characteristics of the aggregated message.

6. The method according to claim 5, characterized in that, The temporal channel output features and spatial channel output features are fused using a concatenation and nonlinear mapping method, and the expression is as follows: , In the formula, Output features for the time channel; F represents the spatial channel output feature; F represents the spatiotemporal fusion feature. The fused features undergo a nonlinear transformation via a fully connected network, the expression of which is: , In the formula, For the final fusion output features; , This is the weight matrix of the fully connected layer; , For bias terms; is the linear rectified activation function; F is the fused feature after splicing.

7. The method according to claim 6, characterized in that, A self-attention weighted aggregation mechanism is used to enhance features at key time steps. Its expression is as follows: , In the formula, The attention weights are for the i-th time step. The aggregated context vector expression is: , In the formula, This is the feature vector corresponding to the i-th time step; The future power prediction result is output through a fully connected network, and its expression is: , In the formula, C is the predicted power sequence; C is the context aggregation vector. This is the weight matrix of the fully connected layer. This is used to output the bias vector of the fully connected layer.

8. A load power prediction system for the steel industry based on a spatiotemporal graph fusion model, characterized in that, include: The data acquisition module is used to acquire multi-dimensional industrial data in the short-process production of the steel industry, including equipment operating status data, historical power sequence data, production task execution data, time-of-use electricity price data, and timestamp data. The heterogeneous graph construction module is used to construct a heterogeneous graph of steel production tasks and equipment. Nodes represent production tasks and production equipment, and edges represent equipment occupancy relationships, process flow relationships, and temporal dependencies. The time feature extraction module is used to construct a time channel based on the Transformer encoder and extract time-series features from the historical power sequence to obtain long-term time-series dependency features in the load sequence. The spatial feature extraction module is used to construct a spatial channel based on a heterogeneous graph convolutional network, model the topological relationship between production tasks and equipment, and realize feature propagation between nodes through a message passing mechanism. The heterogeneous graph includes: Task nodes are used to represent production tasks in the steel production process; Equipment nodes are used to represent electric arc furnace (EAF), argon-oxygen decarburization furnace (AOD), ladle furnace (LF), and continuous casting machine (CC). Device occupancy edge is used to describe the occupancy relationship between tasks and devices; The process flow edge is used to describe the sequential relationship of the production process; Temporal dependency edges are used to describe the temporal continuity between production tasks; The spatiotemporal fusion module is used to fuse the output features of the temporal channel and the output features of the spatial channel, and performs weighted aggregation through a self-attention mechanism to form fused features; The power prediction module is used to input the fused features into the power prediction layer and output the power prediction results of the steel industry load for a future time period.

9. A computer storage medium storing a computer program, characterized in that, When the program is run by the processor, it is able to perform the method described in any one of claims 1-7.

Citation Information

Patent Citations

  • Multi-source heterogeneous energy load prediction method and system for manufacturing enterprise cluster

    CN120782042A

  • Load prediction method based on heterogeneous graph and self-supervised training

    CN121688812A