Iron and steel industry user load prediction method based on physical guidance graph neural network

Abstract

The invention provides an iron and steel industry user load prediction method based on a physical guidance graph neural network. Existing research lacks overall analysis on iron and steel enterprises, the coupling influence of electrical characteristics-physical processes in the iron and steel enterprises is separated, and the change characteristics of substances-processes-loads of the iron and steel enterprises cannot be reflected. Therefore, the invention provides the iron and steel industry user load prediction method based on the physical guidance graph neural network. The method comprises the following steps: firstly, constructing a physical information primary function based on a transmission process mechanism so as to accurately reflect a coupling and matching relationship between connected processes; secondly, the iron and steel enterprise network structure is converted into graph data suitable for the prediction model; and finally, designing a group of graph neural network updating strategies which can take a physical information primary function as a weight updating function in neural network edge updating, so that the prediction model has certain interpretability physically.

Description

Technical Field

[0001] The present invention relates to a method for predicting the load of steel industry users based on a physics-guided graph neural network, and particularly relates to the field of mathematical material modeling. Background Art

[0002] The "double high" characteristics presented by the new power system pose new challenges to the real-time balance of power generation and load changes. Accurate and reliable load forecasting is not only a powerful guarantee for realizing the dynamic balance, safe and stable operation of the power system, but also an important basis for arranging power production and scheduling plans.

[0003] The steel industry belongs to an energy-intensive industry, and the total power demand accounts for a large proportion. Load forecasting is also the basis for enterprises to formulate power supply plans and is of great significance for the economic operation of steel industry users. However, different from traditional regional loads, the characteristics of steel industry loads are significant. It is less affected by climate and mainly affected by production conditions, market demand, etc., showing strong volatility and randomness, which makes load forecasting a major problem faced by steel enterprises.

[0004] With the rapid development of the new generation of artificial intelligence, the use of data-driven models for industrial load forecasting has been widely applied. Such methods mainly include traditional machine learning and deep learning. Among them, traditional machine learning algorithms include extreme learning machines, support vector regression, decision trees, etc. Load forecasting is carried out by constructing a regression model, but it needs to rely on manual feature extraction, and there are problems of incomplete or inaccurate feature extraction.

[0005] In recent years, deep learning methods represented by neural networks in machine learning have achieved better results in power load forecasting due to their powerful data fitting and feature extraction capabilities. However, the steel enterprise has gradually evolved into a dynamic and nonlinear complex system. The traditional neural network model ignores the transfer coupling law between the material flow and energy flow in the system, and it is difficult to meet the demand for accurate prediction only through the mapping of input-output relationships.

[0006] In addition, due to the lack of guidance of physical laws in the training data, only using a data-driven model for calculation, the prediction results often have a certain deviation, and even violate the basic physical laws.

[0007] For the prediction research of the power load in the iron and steel industry, the commonly used methods include mechanism models and data-driven models. The mechanism method obtains the physical information of the equipment by solving high-dimensional non-linear equations. However, the power model established by mechanism-based modeling through theoretical calculations not only makes the production process modeling too complex but also does not fully conform to the actual situation on the spot of iron and steel enterprises. In addition, in recent years, data-driven models established relying on intelligent algorithms have gradually been applied to power characteristic modeling. Such methods do not need to rely too much on process mechanisms and knowledge, improve the modeling accuracy to a certain extent, and are used in engineering practice. However, data-driven models have a high dependence on data and lack sufficient interpretability, which is not conducive to the connection and cooperation between various processes. Summary of the Invention

[0008] For the prediction research of the power load in the iron and steel industry, the commonly used methods include mechanism models and data-driven models. The mechanism method obtains the physical information of the equipment by solving high-dimensional non-linear equations. However, the power model established by mechanism-based modeling through theoretical calculations not only makes the production process modeling too complex but also does not fully conform to the actual situation on the spot of iron and steel enterprises. In addition, in recent years, data-driven models established relying on intelligent algorithms have gradually been applied to power characteristic modeling. Such methods do not need to rely too much on process mechanisms and knowledge, improve the modeling accuracy to a certain extent, and are used in engineering practice. However, data-driven models have a high dependence on data and lack sufficient interpretability, which is not conducive to the connection and cooperation between various processes. In view of this, at present, the operation rules and network topology structure of iron and steel enterprises are mainly characterized in the form of a directed weighted graph. This expression method highly coincides with the graph data structure processed by the Graph Neural Network (GNN). Moreover, compared with the limitations of traditional neural networks in processing linear structure data, the graph neural network can more effectively analyze the non-Euclidean relationships contained in graph data. Therefore, a method for predicting the user load in the iron and steel industry based on a physics-guided graph neural network is proposed.

[0009] This invention application proposes a method for predicting the user load in the iron and steel industry based on a physics-guided graph neural network. Accurate and reliable power load prediction is of great significance for the economic and stable operation of the power system in iron and steel enterprises. However, existing research lacks an overall analysis of iron and steel enterprises, separating the coupled influence of "electrical characteristics - physical processes" in iron and steel enterprises and unable to reflect the changing characteristics of "materials - processes - loads" in iron and steel enterprises. In response to the above problems, a method for predicting the user load in the iron and steel industry based on a physics-guided graph neural network is proposed. First, a physical information basis function based on the mechanism of the transmission process is constructed to accurately reflect the coupling and matching relationship between connected processes. Second, the network structure of the iron and steel enterprise is transformed into graph data suitable for the prediction model. Finally, a set of graph neural network update strategies are designed. This strategy can use the physical information basis function as the weight update function in the neural network edge update, making the prediction model physically interpretable to a certain extent.

[0010] To achieve the above objectives, the technical solution adopted in this invention is as follows:

[0011] A method for predicting the user load in the iron and steel industry based on a physics-guided graph neural network, characterized by including the following steps:

[0012] Step S1: Graph data construction; convert the topological structure of the iron and steel enterprise production process into a graph expression according to the requirements of modeling analysis, and use the physical parameters of the production process and the load power parameters to construct graph features; establish a point set model with production and storage functions, establish a line set model with transportation functions, and obtain the correlation function of the collaborative matching relationship between the point set model and the line set model, and the graph model composed of the point set model and the line set model;

[0013] Step S2: Physical information basis function construction; based on the correlation function of the collaborative relationship between the point set model and the line set model in the iron and steel enterprise production process network, construct the physical information basis function of the PGNN model and simulate the coupling process of energy flow and material flow between processes;

[0014] Step S3: Graph network model construction; the PGNN model consists of a graph network layer and a fully connected layer, and the transfer and aggregation of node and edge information can be realized through the update function;

[0015] Step S4: Graph network training and evaluation; during the training process, compare the load prediction value output by the PGNN with the sample label, calculate the loss function, and backpropagate it into the network model to achieve backpropagation. After the training is completed, use the evaluation index to evaluate the prediction performance of the model.

[0016] Further, step S1 includes the following content:

[0017] Step S101: Adopt the EAF process, where the EAF process includes a short process of electric furnace - refining - continuous casting - hot rolling with scrap steel and electricity as the sources.

[0018] Step S102: Analyze the material flow and energy flow of the production process of the steel enterprise.

[0019] Step S103: Establish a point set model.

[0020] Step S104: Establish a line set model.

[0021] Furthermore, establishing the point set model includes the following content:

[0022] In the production process of the steel enterprise, according to the law of conservation of mass and the law of conservation of energy, the point set model should satisfy the following formula constraints.

[0023]

[0024] In the formula, is the raw material item externally supplied or supplied by other processes, which is the main processing object of the point set model; α i represents the auxiliary raw material input in this process; is the raw material item that is recycled from other point set models and added to this point set model to achieve the reuse of resources; represents the waste discharged in this process; represents the material flow entering the next process, which is the intermediate product item or final product item of the point set model; β i is the resource recovery item; represents the energy flow carried by the material flow of the previous process. This part of the energy is mainly the physical heat carried by various material flows into the point set model; is the external power supply of this process; represents the energy composition recovered and self - used from the waste heat and residual energy generated by the point set model; is the energy loss in this process; is the physical heat carried by the product output from the point set model; is the energy recovered and reused in this process.

[0025] Furthermore, establishing the line set model includes the following content:

[0026] According to the ladle structure and related parameters, and processed by one - dimensional steady - state heat conduction, the heat fluxes q 1 of the ladle bottom and slag layer and q 2 of the ladle wall are respectively:

[0027]

[0028] In the formula: T 1 is the inner surface of the ladle; Tf1 is the ambient temperature; δ i is the thickness of the i-th layer of wall; λ i is the thermal conductivity of the i-th layer of wall; α 1 is the radiation heat transfer coefficient; α 2 is the convective heat transfer coefficient; r i is the radius of the i-th layer of ladle wall; r n+1 is the outer radius of the ladle.

[0029] Obtained based on the above formula and the heat conduction equation, within the range of the formula constraints, the temperature drop of the molten steel and the transportation time are linearly related:

[0030] Q loss = c s m·t w ;

[0031] T = kt w + b;

[0032] In the formula: c s is the specific heat capacity of the molten steel; Q loss is the heat loss value of the molten steel during the intermediate transportation of the ladle; k, b are constants; T is the temperature of the molten steel varying with t w .

[0033] The formula for the temperature drop of the molten steel with time is as follows:

[0034] t w = t 1 + t 2 + t 3 + t 4 ;

[0035] In the formula: t 1 represents the transportation time; t 2 represents the simple processing time during transportation; t 3 represents the waiting time, that is, when there is a product in the workstation being normally processed, it needs to wait until it is completed before it can enter; t 4 represents the failure waiting time caused by unexpected events.

[0036] Furthermore, step S2 includes constructing the physical information basis function of the PGNN model based on the correlation function of the collaborative relationship between the point set model and the line set model in the production process network of the iron and steel enterprise, and finally simulating the coupling process of energy flow and material flow between processes; the implementation part of step S2 includes step S201 and step S202, where step S201 includes the following content:

[0037] The atlas model of the production process of the iron and steel enterprise includes two parts: graph expression and graph features. The mathematical description of its atlas model G is shown in the following formula:

[0038] G = {N, E, U};

[0039] Among them: N is the node feature, used to represent each process in the production process, including smelting, refining, continuous casting, and rolling. N = {n i}, where i = 1:N v represents the set of N v nodes in the figure, and n i is the vector representation of node i, used to express electrical physical parameters; E is the edge feature, E = {e ij}, where i, j = 1:N v , and e ij is the vector representation of the edge relationship between node i and node j, used to characterize the coupling mechanism between the front and back processes; U is the global feature;

[0040] Furthermore, select the physical parameters of the i-th device as the node feature; for the iron and steel industry users with N v devices, to make the node features of the graph nodes n i corresponding to each device have the same dimension, the features of the graph node n i are defined as follows:

[0041] n i = {m i , T i , P i};

[0042] In the formula: m i is the mass of the material input in process i; T i is the temperature of the material input in process i; P i is the active power during the operation of process i;

[0043] E is the edge feature, E = {e ij}, where the mathematical expression of the three-dimensional vector e ij is as follows:

[0044] e ij = {t ij , m ij , l ij};

[0045] Among them, t ij is the transportation time, m ij is the transportation mass, and l ij is the transportation path.

[0046] Furthermore, the implementation part of step S2 includes step S201 and step S202, where step S202 includes the following content:

[0047] Step S202: Introduce the attenuation law. When constructing the physical information basis function, only the temperature loss part during intermediate transportation needs to be considered, as shown in the following formula:

[0048] ΔL = kt w + b;

[0049] where: k and b are constants; ΔL is the molten steel temperature varying with tw.

[0050] Furthermore, Step S3 includes building a graph network model. The PGNN model includes a graph network layer and a fully connected layer. The PGNN model realizes the transmission and aggregation of node and edge information through an update function, specifically as the following steps:

[0051] Step S301: The PGNN model regards equipment and processes as nodes of the graph network model, regards the influence relationship between processes as the edges of the graph network model, and represents the short process of the steel enterprise production process with a graph network model; and updates the edges, nodes, and global features of the steel enterprise production process graph network model by sending and updating messages between the nodes of the steel enterprise production process graph network model; furthermore, the PGNN model transforms the input graph network model into a set of node embedding vectors; and then uses the node embedding vectors as inputs through the fully connected layer to predict the output power of each equipment or process, so as to predict the total load of the entire steel enterprise; among them, in the message transmission and update process, the update function and process introduce a process attenuation mechanism transmission model and use this transmission model as the weight update function in message transmission;

[0052] The input data of the PGNN model includes the historical active power data, process flow data, and transportation link information of each process in the steel enterprise production process as the model input data set; among them, the process flow data includes the start-stop time sequence of equipment and the topology of the enterprise production process, and the transportation link data covers ladle parameters and the transportation time between processes; the output data is the predicted data of the total active power of the steel enterprise;

[0053] Step S302: Define the update function in the graph network model, and the specific content is as follows:

[0054] Define the GN layer as the graph network layer in the graph neural network. The GN layer is responsible for processing graph structure data and performing information propagation and feature aggregation between nodes; among them, the GN layer uses the node update function φ n 、edge update function φ e 、global update function φ g 、weight update function φ w to complete the update of information during the message transmission process; each update function respectively includes its own node aggregation function ρ n 、edge aggregation function ρ e 、global aggregation function ρ g 、weight aggregation function ρw ; where n, e, g, and w represent node features, edge features, global features, and weight features respectively;

[0055] Step S303: During the message passing process, the GN layer of the information passing network uses the update functions φe, φn, φg, and φw of the above-defined formula to complete the information update; when a GN layer receives an input of a graph, the calculation usually proceeds from the edges to the nodes and then to the global level.

[0056] Furthermore, the node update function φ of the GN layer n , edge update function φ e , global update function φ g , and weight update function φ w also include the following:

[0057] 1) Edge update: The edge update function φ e uses the sender node feature n j , the receiver node feature n i , and the edge feature e ij between the two nodes to generate the updated edge feature e ij '; The GN layer uses the weight calculation function φ w to dynamically adjust the weight according to the material processing and transportation conditions of the internal production processes of the iron and steel enterprise; and selects the temperature value and quality of the material processed by the equipment as the input of φ w , φ w calculates the weight W ij for obtaining the updated edge e ij '; Using the influence relationship in the above mechanism model to obtain the weight to adjust the influence intensity between equipment or processes, this calculation is physically meaningful and has strong interpretability;

[0058] 2) Node update: The node update function φ n uses the current input node feature n i , the updated edge feature e ij ', and the global feature g to update the node feature; The GN layer uses the edge aggregation function ρ e to reduce the edge feature set to a single aggregated edge feature e ij '; The node update function φ n processes the aggregated edge feature e ij ' and the current node feature n i to generate the updated node feature n i ';

[0059] 3) Global update: The global update function φ gUsed to update the global features. The GN layer uses the updated edge and node features to update the global features. For this purpose, the GN layer utilizes the aggregation functions ρ n and ρ g to aggregate all edge features E' and all node features N' into two vectors e' and n';

[0060] 4) Weight update: The edge weight w ij is updated, representing the attenuation degree of the intermediate transportation process. The weight update function φ w is used for updating the edge information between nodes in the graph neural network, and the expression is the physical information basis function constructed in step S2.

[0061] Furthermore, step S4 includes the following content:

[0062] Step S4 includes graph network training and evaluation. Compare the load prediction value output by the PGNN with the sample label, calculate the loss function, and backpropagate it into the network model to achieve backpropagation. After training is completed, use the evaluation index to evaluate the prediction performance of the model. Specifically, it is the following steps:

[0063] Step S401: The graph network training is as follows:

[0064] The training steps of the PGNN model are as follows: First, randomly divide the collected data into a training set, a test set, and a validation set; Second, standardize the sample data of the steel enterprise;

[0065]

[0066] In the formula: x is the value of the original data; u is the mean of the original data; σ is the standard deviation of the original data set. The standardized value z represents the deviation degree of the original data relative to the mean. By comparing and analyzing the z values, the influence of data outliers on the analysis results is reduced.

[0067] Finally, use the trained model to predict the active power of the steel enterprise load. Use the feature sets of the steel enterprise point set model and the line set model as the input to establish the PGNN graph network model to predict the steel enterprise power load for the next T time steps, so that the error between the predicted load and the actual load is minimized;

[0068] Step S402: The evaluation method of this invention application adopts the following evaluation indexes: Mean Absolute Error (MAE) X MAE 、Mean Absolute Percentage Error (MAPE) X MAPE and Root Mean Square Error (RMSE) XRMSE Quantitatively evaluate the simulation effect of the model power using indicators such as, and its calculation method is as follows:

[0069]

[0070]

[0071] In the formula: y i represents the fitting value of the model at the i-th moment; represents the measured value of the power at the i-th moment; N is the number of data.

[0072] The present invention has the following advantages:

[0073] 1) A user load prediction method for the iron and steel industry based on a physics-guided graph neural network is proposed. Combining the technological process of iron and steel enterprises and considering the coupling relationship of "electrical characteristics - physical process" further improves the accuracy of the power model and is more in line with the actual production process.

[0074] 2) The topological structure of iron and steel enterprises is converted into a graph expression according to the requirements of modeling analysis, and graph features are constructed using production process physical parameters and load power parameters.

[0075] 3) Based on the correlation function of the cooperative relationship between the point set model and the line set model in the process network of iron and steel enterprises, a physical information basis function of the PGNN model is constructed to simulate the coupling process of energy flow and material flow between processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a schematic flow chart of the present invention.

[0077] Figure 2 is a schematic diagram of the short-process system architecture of iron and steel enterprises in the present invention.

[0078] Figure 3 is a schematic diagram of the material flow and energy flow model in the present invention.

[0079] Figure 4 is a schematic diagram of the graph expression of iron and steel enterprises in the present invention.

[0080] Figure 5 is a schematic diagram of the GN layer update strategy in the present invention.

[0081] Figure 6 is a schematic diagram of the model training process of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0082] The following combines the attached Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 ,Figure 6 , the technical solution of the present invention will be specifically described.

[0083] As Figure 1 shown, first, a physical information-based function is constructed based on the mechanism of the transmission process to accurately reflect the coupling and matching relationship between connected processes; second, the network structure of the iron and steel enterprise is transformed into graph data suitable for the prediction model; finally, a set of graph neural network update strategies are designed, and this strategy can use the physical information-based function as the weight update function in the neural network edge update, making the prediction model physically interpretable to a certain extent.

[0084] A method for predicting the user load in the iron and steel industry based on a physics-guided graph neural network, characterized by comprising the following steps:

[0085] Step S1: Graph data construction. The topological structure of the iron and steel enterprise is converted into a graph expression according to the requirements of modeling analysis, and graph features are constructed using the physical parameters of the production process and the load power parameters.

[0086] Step S2: Construction of the physical information-based function. Based on the correlation function of the coordination relationship between the point set model and the line set model in the process network of the iron and steel enterprise, the physical information-based function of the PGNN model is constructed to simulate the coupling process of energy flow and material flow between processes.

[0087] Step S3: Construction of the graph network model. The PGNN model consists of a graph network layer and a fully connected layer, and the transfer and aggregation of node and edge information can be realized through the update function.

[0088] Step S4: Training and evaluation of the graph network. During the training process, the load prediction value output by the PGNN is compared with the sample label, the loss function is calculated, and it is backpropagated into the network model to realize backpropagation. After the training is completed, the evaluation index is used to evaluate the prediction performance of the model.

[0089] Furthermore, in the graph data construction described in step S1, the topological structure of the iron and steel enterprise is converted into a graph expression according to the requirements of modeling analysis, and graph features are constructed using the physical parameters of the production process and the load power parameters. Specifically, it is the following steps:

[0090] Step S101: The development of the modern iron and steel industry has formed two basic processes, the blast furnace-converter-hot rolling-deep processing long process (BF-BOF process) with iron ore and coal as the source and the electric furnace-refining-continuous casting-hot rolling short process (EAF process) with scrap steel and electricity as the source. The chart content of step S101 is as Figure 2 shown.

[0091] At present, with the advancement of China's electrification transformation, the production process of industrial users will mainly be a process of material processing and conversion using electrical energy with the aid of electrical equipment. At the same time, in the future, scrap steel resources and abundant power resources will strongly support the advancement of the all-scrap electric arc furnace (EAF) short process. Developing the EAF process is also the most easily achievable way for the green and low-carbon development of the iron and steel industry. Therefore, this invention will take the EAF process as an example to explore the method of predicting the power load of iron and steel enterprises based on a physics-guided graph neural network.

[0092] Step S102: Analyze the material flow and energy flow of the iron and steel enterprise, and its diagrammatic content is as Figure 3 shown.

[0093] The iron and steel production process is a process of processing scrap steel or iron ore to obtain finished steel products. Academician Yin Ruiyu pointed out that the above production process can be regarded as a complex process network composed of various processes. During the entire production process, the material flow operates dynamically and orderly along a specific "process network" under the drive and action of the energy flow. The iron and steel manufacturing process can be abstractly described through graph data. Since the short process is a production process that uses electricity as the heat source and 100% scrap steel resources as raw materials for steelmaking, generally, the material flow is regarded as molten steel and steel billets, and the energy flow is regarded as the thermal energy and externally supplied electrical energy carried by them.

[0094] In the description of graph data, an iron and steel enterprise can be regarded as a "graph" composed of a set of point models with production or storage functions, a set of line models with transportation functions, and a correlation function describing the collaborative matching relationship between the set of point models and the set of line models. Further, the set of point models is composed of processes or equipment with processing properties. The set of line models is composed of equipment with transportation properties in the system, and due to the limitations of the production process, the transportation direction is often fixed. The operating characteristics of the set of point models and the set of line models will be described in detail below.

[0095] Step S103: The operating characteristic model of the set of point models, and its content is shown in picture form as Figure 3 shown.

[0096] In an iron and steel enterprise, according to the law of conservation of mass and the law of conservation of energy, each set of point models should satisfy the following formula constraints.

[0097]

[0098] In the formula, is the raw material item externally supplied or supplied by other processes, which is the main processing object of the set of point models; α i represents the auxiliary raw material input in this process; is the raw material item that is recycled by other sets of point models and then added to this set of point models to achieve the reuse of resources; represents the waste discharged in this process; Represents the material flow entering the next process, which is an intermediate product item or a final product item of the point set model; β i Is the resource recovery item; Represents the energy flow carried by the material flow of the previous process. This part of the energy is mainly the physical heat carried by various material flows into the point set model; Is the external power supply for this process; Represents the energy composition recovered and self - used from the waste heat and residual energy generated by the point set model; Is the energy lost in this process; Is the physical heat carried by the product output from the point set model; Is the energy recovered and reused in this process.

[0099] Step S104: The operating characteristic model of the line set model is specifically described as follows:

[0100] The line set model also belongs to the coupling process of energy flow and material flow. This application of the present invention focuses on analyzing the attenuation laws of energy and mass on the line set model, that is, the connection laws of physical parameters between processes. On this basis, further analyze the main factors affecting the attenuation rate in the process of energy flow and material flow transportation, and establish relevant analysis models. Currently, the main parameters involved in the calculation of the mechanism model are the amount of matter, temperature, and transportation time.

[0101] Solve sequentially from the three aspects of the amount of matter, temperature, and transportation time:

[0102] For the amount of matter, whether it is the transportation method or the transportation path, the influence on it is limited, so the attenuation change during the transportation process can be ignored.

[0103] For the loss of energy flow, especially the loss of temperature, it generally cannot be ignored. According to the ladle structure and related parameters, treated by one - dimensional steady - state heat conduction, the heat fluxes q 1 of the ladle bottom and the slag layer and q 2 of the ladle wall are respectively:

[0104]

[0105] In the formula: T 1 Is the inner surface of the ladle; T f1 Is the ambient temperature; δ i Is the thickness of the i - th layer of the wall; λ i Is the thermal conductivity of the i - th layer of the wall; α 1 Is the radiation heat transfer coefficient; α 2 Is the convective heat transfer coefficient; r i Is the radius of the i - th layer of the ladle wall; r n+1 Is the outer radius of the ladle.

[0106] Obtained according to the above formula and the heat conduction equation, within the range of the formula constraints, the temperature drop of molten steel and the transportation time are linearly related:

[0107] Q loss = c s m·t w ;

[0108] T = kt w + b;

[0109] Where: c s is the specific heat capacity of molten steel; Q loss is the heat loss value of molten steel during the intermediate transportation of the ladle; k and b are constants; T is the temperature of molten steel that changes with t w .

[0110] During the production process, the scrap steel smelted is transported by the ladle, and once the molten steel is loaded into the ladle, the ladle will not be replaced throughout the process. Therefore, the transportation method is certain during the transportation process, and the temperature drop of molten steel is most significantly affected by the time factor:

[0111] t w = t 1 + t 2 + t 3 + t 4 ;

[0112] Where: t 1 represents the transportation time; t 2 represents the simple processing time during transportation; t 3 represents the waiting time, that is, when there is a product in the workstation being normally processed, it needs to wait until it is completed before it can enter; t 4 represents the failure waiting time caused by unexpected events.

[0113] In summary, there are attenuation laws of temperature and quality in the transportation process of material flow and energy flow between processes, which is the basis for the construction of the physical information basis function in the graph neural network mentioned in step S2 of this invention application.

[0114] Furthermore, the construction of the physical information basis function described in step S2 is based on the correlation function of the collaborative relationship between the point set model and the line set model in the process network of iron and steel enterprises, and constructs the physical information basis function of the PGNN model to simulate the coupling process of energy flow and material flow between processes. Specifically, it is the following steps:

[0115] Step S201: The graph data of iron and steel enterprises includes two parts: graph expression and graph features, and its mathematical description is shown in the following formula.

[0116] G = {N, E, U};

[0117] Where: N is the node feature, used to represent each process in the production process, including smelting, refining, continuous casting, and rolling, N = {n i}, where i = 1:N v represents the set of N v nodes in the figure, and n i is the vector representation of node i, used to express electrical physical parameters; E is the edge feature, E = {e ij}, where i, j = 1:N v , and e ij is the vector representation of the edge relationship between node i and node j, used to characterize the coupling mechanism between the front and back processes; U is the global feature.

[0118] Select the physical parameters of the i-th device as the node feature. For a steel industry user with N v devices, in order to make the node features of the graph nodes n i corresponding to each device have the same dimension, the features of the graph node n i are defined as follows:

[0119] n i = {m i , T i , P i};

[0120] Where: m i is the mass of the material input in process i; T i is the temperature of the material input in process i; P i is the active power during the operation of process i.

[0121] Each edge feature is a three-dimensional vector including the transportation time t ij , the transportation mass m ij , and the transportation path l ij .

[0122] e ij = {t ij , m ij , l ij};

[0123] After converting the production process of the steel enterprise into an edge graph, the edge set E corresponds to the connection relationship between each process. Based on this, the adjacency matrix A of each process of the steel enterprise can be constructed. The adjacency matrix A characterizes the current topological structure of each process of the steel enterprise. If the topology of the production process of the steel enterprise changes, the corresponding adjacency matrix of the graph data also needs to be modified.

[0124] In summary, with the node set N and edge set E corresponding to the steel enterprise, the corresponding graph data G is constructed, as Figure 4 shown.

[0125] Step S202: The purpose of introducing the attenuation law is to describe the internal power change law of the system under the influence of logistics energy flow, and to better consider the coupling and matching relationship between different processes.

[0126] This invention application constructs a physical information basis function based on the correlation function of the collaborative relationship between the point set model and the line set model. The material quantity and temperature after the attenuation of energy flow and logistics are obtained by subtracting the loss during the intermediate transportation link from the initial temperature. Therefore, when constructing the physical information basis function, only the temperature loss part during the intermediate transportation needs to be considered, as shown in the following formula.

[0127] ΔL = kt w + b;

[0128] In the formula: k and b are constants; ΔL is the molten steel temperature that changes with t w variation.

[0129] The above formula will be used as the weight update function in message passing to allocate corresponding weights to each node during the update process, that is, to adjust the "intensity" of the interaction of the attenuation effects received by each wind process, similar to the logistics energy flow coupling effect generated by the upstream process on the downstream process physically.

[0130] Furthermore, in the construction of the graph network model described in step S3, the PGNN model consists of a graph network layer and a fully connected layer, and can realize the transmission and aggregation of node and edge information through the update function, specifically as the following steps:

[0131] Step S301: The input data of the graph network is the historical active power data, process flow data, and transportation link information of each process in the steel enterprise as the model input data set. Among them, the process flow data includes the start-stop time sequence of equipment and the topology of the enterprise production process, and the transportation link data covers ladle parameters and the transportation time between processes; the output data is the predicted data of the total active power of the steel enterprise.

[0132] Step S302: Define the update function in the graph network model as follows:

[0133] The GN layer uses the node update function φ n , edge update function φ e , global update function φ g , weight update function φ w and other 4 functions to complete the information update during the message passing process. Each update function respectively includes its own node aggregation function ρ n , edge aggregation function ρ e , global aggregation function ρ g , weight aggregation function ρ w . Among them, n, e, g, and w represent node features, edge features, global features, and weight features respectively.

[0134] 1) Edge update: Edge update function φ e Utilize the sender node feature n j , the receiver node feature n i and the feature e of the edge between the two nodes ij to generate the updated edge feature e ij '. The GN layer uses the weight calculation function φ w to dynamically adjust the weight according to the material processing and transportation conditions of the internal production processes of the iron and steel enterprise. And select the temperature value and quality of the material processed by the equipment as the input of φ w , φ w calculate the weight W ij , which is used to obtain the updated edge e ij '. Use the influence relationship in the above mechanism model to obtain the weight to adjust the influence strength between equipment or processes. This calculation is physically meaningful and has strong interpretability.

[0135] 2) Node update: Node update function φ n Use the current input node feature n i , the updated edge feature e ij ' and the global feature g to update the node feature. The GN layer uses the edge aggregation function ρ e to reduce the set of edge features to a single aggregated edge feature e ij '. The node update function φ n processes the aggregated edge feature e ij ' and the current node feature n i to generate the updated node feature n i '.

[0136] 3) Global update: Global update function φ g used to update the global feature. The GN layer uses the updated edge and node features to update the global feature. For this purpose, the GN layer uses the aggregation functions ρ n and ρ g to aggregate all edge features E' and all node features N' into two vectors e' and n'.

[0137] 4) Weight update: Edge weight w ij is updated, which represents the attenuation degree of the intermediate transportation process. The weight update function φ w is used for the update of the edge information between nodes in the graph neural network, and the expression is the physical information basis function constructed in step S202 above.

[0138] Step S303: During the message passing process, the GN layer of the information passing network uses the update functions φ e , φ n, φ g and φ w to complete the information update.

[0139] When a GN layer receives the input of a graph, the calculation usually proceeds from the edges to the nodes and then to the global. As Figure 5 shows the update process of a GN layer.

[0140] Furthermore, for the graph network training and evaluation described in step S4, during the training process, the load prediction value output by the PGNN is compared with the sample label, the loss function is calculated, and it is backpropagated into the network model to achieve backpropagation. After the training is completed, the evaluation metrics are used to evaluate the prediction performance of the model. Specifically, the following steps are as follows:

[0141] Step S401: The specific description of the graph network training is as follows:

[0142] The training steps of the PGNN model are specifically as Figure 6 shown. First, the collected data is randomly divided into a training set, a test set, and a validation set. Secondly, the sample data of steel enterprises contains many electrical and non-electrical quantities, with inconsistent dimensions and large differences in numerical ranges. Therefore, data standardization is required.

[0143]

[0144] In the formula: x is the value of the original data; u is the mean of the original data; σ is the standard deviation of the original data set. The standardized value z represents the deviation degree of the original data relative to the mean. By comparing and analyzing the z values, the influence of data outliers on the analysis results is reduced.

[0145] Finally, the trained model is used to predict the active power of steel enterprises. The feature sets of the enterprise point set model and the line set model are used as the input for establishing the PGNN graph network model to predict the power load of steel enterprises for the next T time steps, so as to minimize the error between the predicted load and the actual load.

[0146] Step S402: The evaluation method of this invention application adopts the following evaluation metrics: Mean Absolute Error (MAE) X MAE , Mean Absolute Percentage Error (MAPE) X MAPE and Root Mean Square Error (RMSE) X RMSE and other metrics to quantitatively evaluate the power simulation effect of the model. The calculation methods are as follows:

[0147]

[0148] where: y i represents the fitted value of the model at time i; represents the measured value of power at time i; N is the number of data.

[0149] In the present invention, the topological structure of the iron and steel enterprise is converted into a graph expression according to the requirements of modeling analysis, and the physical parameters of the production process and the load power parameters are used to construct the graph features; based on the correlation function of the coordination relationship between "points" and "lines" in the process network of the iron and steel enterprise, the physical information basis function of the PGNN model is constructed to simulate the coupling process of energy flow and material flow between processes. Therefore, the solution of the present invention is typical.

[0150] The above are the preferred embodiments of the present invention. All changes made according to the technical solution of the present invention, when the functions and effects produced do not exceed the scope of the technical solution of the present invention, fall within the protection scope of the present invention.

Claims

1. A method for forecasting user load in the steel industry based on a physical guided graph neural network, characterized in that: The following steps are involved: Step S1: Graph data construction; convert the topological structure of the steel enterprise production process into a graph expression according to the modeling and analysis requirements, and use the physical parameters of the production process and the load power parameters to construct the graph features; establish a point set model with production and storage functions, establish a line set model with transportation functions, and obtain the correlation function of the collaborative matching relationship between the point set model and the line set model, and the graph model composed of the point set model and the line set model; Step S2: Construction of physical information basis function: Based on the correlation function of the collaborative relationship between the point set model and the line set model in the production process network of the steel enterprise, the physical information basis function of the PGNN model is constructed, and the coupling process of energy flow and logistics between processes is simulated; Step S3: Build a graph network model; the PGNN model includes a graph network layer and a fully connected layer. The PGNN model realizes the transmission and aggregation of node and edge information through an update function; Step S4: graph network training and evaluation; During the training process, the load prediction value output by PGNN is compared with the sample label, the loss function is calculated, and then sent back to the network model for back propagation. After the training is completed, the evaluation index is used to evaluate the model prediction performance.

2. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 1 is characterized in that: Step S1 includes the following contents: Step S101: adopting the EAF process, wherein the EAF process includes a short process of electric furnace-refining-continuous casting-hot rolling with scrap steel and electricity as the source; Step S102: analyzing the material flow and energy flow of the steel enterprise's production process; Step S103: establishing a point set model; Step S104: Establish a line set model.

3. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 2 is characterized in that: Building a point set model includes the following: In the production process of steel enterprises, according to the law of conservation of mass and energy, the point set model should satisfy the following constraints: In the formula, It is the raw material supplied from outside or from other processes, and is the main processing object of the point set model; α i Indicates the auxiliary raw materials input into this process; After being recycled by other point set models, the raw material items of this point set model are added to realize the reuse of resources; Indicates that waste is discharged from this process; Indicates the material flow entering the next process, which is the intermediate product item or final product item of the point set model; β i It is a resource recovery item; Represents the energy flow carried by the material flow of the previous process. This part of energy is mainly the physical heat carried by various logistics into the point set model; Provides external power for this process; It represents the energy composition of the waste heat and waste energy generated by the point set model for self-use recovery; is the energy lost in this process; is the physical heat carried by the product output by the point set model; Energy recovered and reused in this process.

4. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 2 is characterized in that: Building a line set model includes the following: According to the ladle structure and related parameters, according to the one-dimensional steady-state heat conduction treatment, the heat fluxes of the ladle bottom and slag layer q1 and the ladle wall q2 are: Where: T1 is the inner surface of the ladle; T f1 is the ambient temperature; i is the thickness of the i-th wall; λ i is the thermal conductivity of the i-th wall; α1 is the radiation heat transfer coefficient; α2 is the convection heat transfer coefficient; r i is the radius of the i-th layer wall; r n+1 is the outer radius of the ladle; According to the above formula and heat conduction equation, within the range of the constraint, the temperature drop of molten steel and the transportation time are linearly related: Q loss =c s m·t w ; T=kt w +b; Where: c s is the specific heat of molten steel; Q loss is the heat loss value of molten steel during the intermediate transportation of the ladle; k and b are constants; T is the w Changing molten steel temperature; The formula for the temperature drop of molten steel over time is as follows: t w =t1+t2+t3+t4; In the formula: t1 represents the conveying time; t2 represents the simple processing time during the conveying process; t3 represents the waiting time, that is, the workstation has products that are processed normally and needs to wait for completion before entering; t4 represents the waiting time for failures caused by emergencies.

5. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 1 is characterized in that: Step S2 includes constructing the physical information basis function of the PGNN model based on the correlation function of the synergistic relationship between the point set model and the line set model in the production process network of the steel enterprise, and finally simulating the coupling process of energy flow and logistics between processes; the implementation part of step S2 includes step S201 and step S202, wherein step S201 includes the following contents: The atlas model of the steel enterprise production process consists of two parts: graph expression and graph features. The mathematical description of the atlas model G is shown in the following formula: G = {N, E, U}; Where: N is the node feature, which is used to characterize each process in the production process, including smelting, refining, continuous casting and steel rolling, N = {n i }, where i = 1:N v Represents N in the figure v The set of nodes, n i is the vector representation of node i, used to express electrical physical parameters; E is the edge feature, E = {e ij }, where i,j=1:N v , e ij is the vector representation of the edge relationship between node i and node j, which is used to characterize the coupling mechanism between the previous and next processes; U is the global feature; Furthermore, the physical parameters of the i-th device are selected as the characteristics of the node; for N v Steel industry users with 1000 devices, in order to make the graph node n i The corresponding node feature dimensions are the same, and the graph node n i The feature definition is as follows: n i ={m i ,T i ,P i }; Where: m i is the mass of the material when it is put into process i; T i is the temperature of the material when it is put into process i; P i is the active power of process i during operation; E is the edge feature, E={e ij }, where the three-dimensional vector e ij The mathematical expression is as follows: e ij ={t ij ,m ij ,l ij }; where t ij is the transportation time, m ij is the transport quality, l ij is the transport path.

6. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 1 is characterized in that: The implementation part of step S2 includes step S201 and step S202, wherein step S202 includes the following contents: Step S202: Introducing the attenuation law, when constructing the physical information basis function, only the temperature loss part of the intermediate transportation needs to be considered, as shown in the following formula: ΔL=kt w +b; Among them: k, b are constants; ΔL is the molten steel temperature that changes with tw.

7. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 1 is characterized in that: Step S3 includes building a graph network model, wherein the PGNN model includes a graph network layer and a fully connected layer. The PGNN model implements the transmission and aggregation of node and edge information through an update function, including the following steps: Step S301: The PGNN model includes taking equipment and processes as nodes of a graph network model, taking the influence relationship between processes as edges of the graph network model, and representing the short process in the production process of a steel enterprise with a graph network model; and updating the edges, nodes and global features of the graph network model of the steel enterprise production process by sending and updating messages between the nodes of the graph network model of the steel enterprise production process; further, the PGNN model converts the input graph network model into a set of node embedding vectors; and then uses the node embedding vectors as input through a fully connected layer to predict the output power of each device or process, thereby predicting the total load of the entire steel enterprise; wherein, in the process of message transmission and update, the update function and process introduce a process attenuation mechanism transmission model, and use this transmission model as a weight update function in message transmission; The input data of the PGNN model include historical data of active power of each process of the steel enterprise production process, process data and transportation link information as the model input data set; the process data includes the start and stop sequence of the equipment and the enterprise production process topology, and the transportation link data covers the ladle parameters and the transportation time between processes; the output data is the predicted data of the total active power of the steel enterprise; Step S302: define the update function in the graph network model, the specific contents are as follows: The GN layer is defined as the graph network layer in the graph neural network. The GN layer is responsible for processing graph structure data, performing information propagation and feature aggregation between nodes. The GN layer uses the node update function φ n , edge update function φ e , global update function φ g , weight update function φ w The information is updated during the message passing process; each update function contains its own node aggregation function ρ n , edge aggregation function ρ e , global aggregation function ρ g , weight aggregation function ρ w ; Among them, n, e, g, and w represent node features, edge features, global features, and weight features, respectively; Step S303: During the message passing process, the GN layer of the information passing network uses the update functions φe, φn, φg and φw defined above to complete the information update; when a GN layer receives a graph input, the calculation is usually performed from the edge to the node and then to the global.

8. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 7 is characterized in that: The node update function φ of the GN layer n , edge update function φ e , global update function φ g , weight update function φ w Also included are the following: 1) Edge update: edge update function φ e Using the sender node feature n j , receiving node characteristics n i and the feature e of the edge between two nodes ij Generate the updated edge feature e ij '; GN layer uses weight calculation function φ w The weight is dynamically adjusted according to the material processing and transportation conditions of the internal production process of the steel enterprise; and the temperature value and quality of the material processed by the equipment are selected as φ w Input, φ w Calculate the weight W ij , used to get the updated edge e ij '; Using the influence relationship in the above mechanism model to obtain weights to adjust the intensity of the influence between equipment or processes, this calculation is physically meaningful and has strong interpretability; 2) Node update: Node update function φ n Use the current input node feature n i , updated edge feature e ij ' and global feature g to update node features; GN layer uses edge aggregation function ρ e Reduce the set of edge features to a single aggregated edge feature e ij '; Node update function φ n Processing the edge features after aggregation ij ' and the current node feature n i To generate the updated node feature n i '; 3) Global update: global update function φ g Used to update global features; the GN layer uses the updated edge and node features to update global features; Among them, the GN layer uses the aggregation function ρ n and ρ g Aggregate all edge features E' and all node features N' into two vectors e' and n'; 4) Weight update: edge weight w ij Update, characterizes the attenuation degree of the intermediate transportation process, weight update function φ w Used to update the edge information between nodes in the graph neural network, the expression is the physical information basis function constructed in step S2.

9. The method for predicting user load in the steel industry based on physical guided graph neural network according to claim 1, characterized in that: Step S4 includes the following contents: Step S4 includes graph network training and evaluation. The load prediction value output by PGNN is compared with the sample label, the loss function is calculated, and the value is sent back to the network model for back propagation. After the training is completed, the model prediction performance is evaluated using the evaluation index. The specific steps are as follows: Step S401: Graph network training is as follows: The steps of PGNN model training are as follows: first, the collected data is randomly divided into training set, test set and validation set; second, the sample data of steel enterprise production process is standardized; In the formula: x is the value of the original data; u is the mean of the original data; σ is the standard deviation of the original data set; the standardized value z represents the degree of deviation of the original data from the mean. By comparing and analyzing the z value, the influence of data outliers on the analysis results can be reduced; Finally, the trained model is used to predict the active power of the steel enterprise load, and the feature set of the steel enterprise point set model and line set model is used as the input to establish the PGNN graph network model to predict the power load of the steel enterprise in the next T time steps, thereby reducing the error between the predicted load and the actual load. Step S402: The evaluation method of this patent adopts the following evaluation indicators: mean absolute error X MAE , mean absolute percentage error X MAPE and the root mean square error X RMSE The power simulation effect of the model is quantitatively evaluated by using indicators such as: Where: y i represents the fitted value of the model at time i; represents the measured value of power at time i; N is the number of data.

Images