LSTM-based ore blending molten iron stable output time sequence prediction system and method
By using an LSTM-based time-series prediction system, combined with multi-source data and dynamic calibration, the mismatch between the stable iron production time and the steel production plan was solved, achieving high-precision prediction and supply-demand matching, and reducing production costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-15
- Publication Date
- 2026-03-27
AI Technical Summary
In existing technologies, the stable production time of molten iron does not match the steel production plan, resulting in high-quality molten iron flowing to low-end steel grades, causing waste of production costs. Furthermore, the lack of a real-time dynamic calibration mechanism and closed-loop simulation verification makes it difficult to capture nonlinear dynamics and hysteresis effects, leading to large prediction errors.
An LSTM-based time series prediction system is adopted. Multi-source time series data is acquired through a data acquisition module. A multi-layer LSTM network combined with an attention mechanism is used to capture long-term data dependencies. A dynamic calibration module corrects the prediction results in real time. Closed-loop simulation verifies the simulated supply and demand matching degree and provides optimization suggestions.
It achieves high-precision prediction of stable molten iron production time, dynamically adapts to production fluctuations, ensures supply and demand matching, reduces prediction errors, provides comprehensive decision support and risk monitoring, and guarantees stable production.
Smart Images
Figure CN121744934A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent scheduling technology for steel production, and in particular to a time-series prediction system and method for stable iron production from ore blending based on LSTM. Background Technology
[0002] In the production of high-quality steel, different steel grades require different grades of high-quality molten iron, and some of this molten iron needs to be supplied to the blast furnace through a special high-quality steel blend. Currently, the response requirements of different steel grades to different grades of molten iron may not match. For example, the stable output time of molten iron in the ironmaking plant often starts earlier than the planned start time of tire cord and bead steel production, and ends later than the planned end time. This results in some high-quality molten iron flowing to low-end steel grades such as building materials and medium-to-high carbon steel, leading to wasted production costs.
[0003] Existing solutions largely rely on manual experience to estimate the stable iron production time, which has the following shortcomings: First, manual estimation does not fully integrate multi-source time-series data such as ore blending composition and blast furnace conditions, making it difficult to capture the nonlinear dynamics and lag effects in the iron production process, resulting in large estimation errors (typically >4 hours). Second, there is a lack of a real-time dynamic calibration mechanism; when there is a delay in the replacement of the mixing bin or fluctuations in the blast furnace temperature, the time estimation results cannot be corrected in time. Third, there is no closed-loop simulation verification step; the production schedule has not been verified for the matching degree of iron supply and demand, which easily leads to supply and demand imbalances. Therefore, there is an urgent need for a data-driven, dynamic optimization, and simulation verification integrated technical solution to solve the problem of mismatch between the stable iron production time and the production plan. Summary of the Invention
[0004] In view of this, the purpose of this invention is to propose a time-series prediction system and method for stable iron production from ore blending based on LSTM, so as to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, this invention provides a time-series prediction system for stable iron production from ore blending based on LSTM, comprising:
[0006] The data acquisition module is used to collect multi-source time-series data on the chemical composition of blended ore, blast furnace condition parameters, and hot metal production indicators, and to preprocess the collected data.
[0007] The LSTM time series prediction module, based on preprocessed multi-source time series data, captures long-term data dependencies through a multi-layer LSTM network structure, and combines an attention mechanism to enhance feature extraction for a specified process period, outputting a stable time window for molten iron to meet standards.
[0008] The dynamic calibration module acquires real-time production data from the front end through a real-time data interface, compares the deviation between the actual output of molten iron and the predicted time window, and corrects the output of the LSTM timing prediction module online.
[0009] The closed-loop simulation verification module is used to receive the time window output by the LSTM timing prediction module or the dynamically calibrated time window, simulate the steel production scheduling plan, calculate the iron supply and demand matching degree, and verify the feasibility of the scheduling plan.
[0010] Preferably, the multi-source time-series data collected by the data acquisition module includes the chemical composition data of the blended ore, which includes total iron content, silicon dioxide content, titanium content, and phosphorus content; the blast furnace condition parameter data includes furnace top temperature, blast temperature, blast pressure, oxygen enrichment rate, and pulverized coal injection rate; and the molten iron production index data includes molten iron temperature, molten iron titanium content, and molten iron phosphorus content.
[0011] The preprocessing includes data cleaning, missing value imputation, and standardization.
[0012] Preferably, the network structure of the LSTM time series prediction module includes an input layer, multiple LSTM hidden layers, an attention layer, and a fully connected output layer;
[0013] The input layer converts the preprocessed multi-source time-series data into a fixed-dimensional feature vector.
[0014] In the multi-layer LSTM hidden layer, the input gate, forget gate and output gate of each LSTM unit adopt the Sigmoid activation function, the cell state update adopts the Tanh activation function, and a Dropout layer is set between adjacent LSTM layers.
[0015] The attention layer enhances the feature contribution of a specified process period by calculating the weight coefficients of the features at each time step.
[0016] The fully connected output layer uses a linear activation function to output the start and end times of stable molten iron production.
[0017] Preferably, the training process of the LSTM time series prediction module includes:
[0018] S1. Construct a training dataset, using complete time series sequences from historical production data as samples. Each sample contains multi-source time series data features and corresponding stable iron production time labels.
[0019] S2. Initialize LSTM network parameters;
[0020] S3. The Adam optimizer is used to minimize the mean square error loss function, which is defined as the mean of the squared deviations between the prediction time window and the actual time window.
[0021] S4. Prevent overfitting of the model by using early stopping. Stop training when the exit condition is met and save the optimal model parameters.
[0022] Preferably, the calibration logic of the dynamic calibration module includes:
[0023] T1. Real-time acquisition of signals indicating completion of material replacement in the mixing bin, sinter composition compliance, and blast furnace temperature fluctuations;
[0024] T2. Calculate the deviation between the actual signal trigger time and the LSTM prediction time. If the deviation is greater than the set value, start the calibration mechanism.
[0025] T3. Based on the deviation value, the parameters of the fully connected layer of the LSTM network are fine-tuned using the gradient descent method to correct the stable iron production time window of the subsequent output.
[0026] Preferably, the produced molten iron includes first-type molten iron and second-type molten iron, and the simulation rules of the closed-loop simulation verification module include:
[0027] U1, Receive the time window output by the LSTM timing prediction module or the dynamically calibrated time window, and set the simulated start time for steel production scheduling;
[0028] U2. Using a set duration as the time unit, calculate the actual output of molten iron, the steelmaking consumption demand, and the molten iron inventory within that time unit.
[0029] U3. Establish priority rules for the use of Class I and Class II molten iron for different steel grades:
[0030] U4. Calculate the iron supply and demand gap rate during the simulation cycle, determine whether the production scheduling plan is feasible, and output optimization suggestions.
[0031] Preferably, the system also includes a visualization and interaction module, including a data monitoring dashboard, a prediction results panel, and a simulation report interface;
[0032] The data monitoring dashboard displays time-series curves of chemical composition of blended ore, blast furnace condition parameters, and hot metal production indicators in real time; the prediction results panel marks the stable hot metal production time window and dynamic calibration records in the form of a time axis; the simulation report interface outputs a hot metal supply and demand matching bar chart, shortage rate statistics, and production scheduling optimization suggestions.
[0033] Preferably, the system further includes a risk management module, which includes:
[0034] The composition monitoring unit is used to collect the composition of primary sinter and molten iron, and triggers an alarm when the composition exceeds the standard.
[0035] The furnace temperature monitoring unit is used to collect blast furnace temperature data in real time. When the furnace temperature fluctuation exceeds the set condition, it pushes furnace condition adjustment suggestions to the ironmaking dispatch terminal.
[0036] The data quality monitoring unit detects and addresses data acquisition distortion or delays in the water tank. When the data loss rate exceeds a set value, it switches to a backup data acquisition channel.
[0037] This invention also provides a time-series prediction method for stable iron production from ore blending based on LSTM, comprising:
[0038] Step 1: Collect multi-source time-series data on chemical composition of blended ore, blast furnace condition parameters, and hot metal production indicators, and perform preprocessing.
[0039] Step 2: Process the preprocessed data through a multi-layer LSTM network and attention mechanism to output the stable production time window of molten iron with ore blending;
[0040] Step 3: Compare the deviation between the real-time production signal and the predicted time, fine-tune the model parameters, and correct the time window;
[0041] Step 4: Based on the obtained time window closed-loop simulation verification module, simulate the production scheduling plan, calculate the supply and demand matching degree, and output feasible solutions or optimization suggestions.
[0042] Preferably, in step 2, the importance of features processed by the multilayer LSTM network and the attention mechanism is ranked as follows: titanium content in the blended ore > blast furnace blast temperature > molten iron temperature > phosphorus content in the blended ore > pulverized coal injection rate.
[0043] The beneficial effects of this invention are:
[0044] 1. High prediction accuracy: By using a multi-layer LSTM network combined with the Attention mechanism and fusing multi-source time series data, the nonlinear dynamics of molten iron production are captured, and the average absolute error of the prediction time window is ≤1.5 hours, which is significantly better than the accuracy of manual prediction.
[0045] 2. Strong dynamic adaptability: Through the dynamic calibration module, time deviations caused by production fluctuations are corrected in real time to ensure that the time window matches the actual production.
[0046] 3. Comprehensive decision support: The closed-loop simulation verification module simulates production scheduling plans, outputs supply and demand matching degree and optimization suggestions, and avoids blind production scheduling;
[0047] 4. Controllable risks: The risk management module monitors the composition, furnace temperature, and data quality in real time, provides timely warnings and solutions, and ensures stable production. Attached Figure Description
[0048] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0049] Figure 1This is a block diagram of the LSTM-based stable iron production time series prediction system according to an embodiment of the present invention.
[0050] Figure 2 This is a schematic diagram of the time series prediction method for stable iron production based on LSTM according to an embodiment of the present invention. Detailed Implementation
[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0052] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0053] Example 1:
[0054] like Figure 1 As shown in the embodiments of this specification, a time series prediction system for stable iron production based on LSTM is provided, including a data acquisition module, an LSTM time series prediction module, a dynamic calibration module, and a closed-loop simulation verification module.
[0055] The data acquisition module is used to collect multi-source time-series data of the entire blast furnace production process and to preprocess the collected data. The multi-source time-series data includes:
[0056] Chemical composition data of blended ore: Tfe (total iron), SiO2 (silicon dioxide), Ti (titanium), and P (phosphorus) content;
[0057] Blast furnace condition parameters: furnace top temperature, blast temperature, blast pressure, oxygen enrichment rate, and pulverized coal injection rate. The data is collected every 10 minutes and the data source is the blast furnace control system.
[0058] Hot metal production indicators include: hot metal temperature, hot metal Ti content, hot metal P content, and hot metal output. The data is collected once per heat (approximately 30-60 minutes per heat) and comes from the performance management module of Yonggang's production scheduling and control system.
[0059] Preprocessing includes:
[0060] Data cleaning: Remove obvious outliers (such as molten iron temperature >1650℃ or <1450℃);
[0061] Missing value imputation: Missing data is imputed using linear interpolation of adjacent time series data to ensure data continuity;
[0062] Standardization: The data is converted into standardized data with a mean of 0 and a standard deviation of 1 using the Z-score formula to eliminate the influence of dimensions.
[0063] The LSTM timing prediction module adopts an end-to-end multi-layer LSTM network structure, combined with the Attention mechanism, to achieve accurate prediction of the stable iron production time window.
[0064] Network architecture design
[0065] Input layer: Convert the preprocessed multi-source time series data into a feature vector with dimensions of (time step, number of features), where the time step is set to 24 (corresponding to 12 hours of data) and the number of features is 12 (covering 4 components of the blended ore, 5 furnace condition parameters of the blast furnace, and 3 indicators of molten iron).
[0066] Hidden layers: Set up 2-3 layers of LSTM units, with 64-128 units per layer. Taking the first layer of LSTM units as an example, its input gates... Forgotten Gate Output gate and cell state The calculation method is as follows:
[0067]
[0068] in, Input features at time t, Let W be the hidden state at time t-1, W be the weight matrix, and b be the bias term. It is the Sigmoid activation function. This is for element-wise multiplication. A Dropout layer is placed between adjacent LSTM layers with a Dropout probability of 0.25 to prevent overfitting.
[0069] Attention layer: Calculates the hidden state at each time step. Weighting coefficients The formula is ,in ( Here is the Attention weight matrix. (This is the bias term), where T is the total number of time steps. The output of the Attention layer is obtained by weighted summation. Strengthen the characteristic contributions of key process periods such as material replacement in the mixing bin and sinter composition compliance;
[0070] Output layer: A fully connected layer is used to... The mapping is to two output values, corresponding to the start time and end time of stable molten iron production, respectively, and the activation function is a linear activation function.
[0071] Model training process
[0072] Dataset construction: Using the complete time series of "ore blending start-up - stable iron production" in historical production as samples, a total of 1000-1500 samples were collected and divided into training set, validation set and test set in a ratio of 7:2:1;
[0073] Parameter initialization: learning rate is set to 0.003, batch size is set to 64, and training epochs are set to 80;
[0074] Loss function and optimizer: The mean squared error (MSE) is used as the loss function, and the Adam optimizer is selected as the optimizer;
[0075] Model optimization: An early stopping method is adopted. Training is stopped and the optimal model parameters are saved when the validation set loss shows no decrease for eight consecutive rounds. Testing shows that the model's mean absolute error within the prediction time window is ≤1.5 hours, meeting production accuracy requirements.
[0076] Dynamic calibration module:
[0077] Due to interference factors such as delayed replacement of mixing bins and fluctuations in blast furnace temperature during actual production, the prediction results need to be corrected in real time through a dynamic calibration module to ensure the accuracy of the time window.
[0078] Calibration trigger conditions
[0079] Two key real-time signals are collected: one is the signal that the mixing bin replacement is complete (triggered when the ironmaking MES system detects that the composition of the mixing material meets P≤0.076 and Ti≤0.105); the other is the signal that the sinter composition meets the standard (triggered 7 hours after the mixing material meets the standard, when the sinter composition meets P≤0.076 and Ti≤0.105).
[0080] Calculate the deviation between the real-time signal trigger time and the LSTM prediction time. If the deviation is greater than 2 hours, start dynamic calibration.
[0081] calibration mechanism
[0082] Parameter fine-tuning: Based on the deviation value, the weight matrix and bias terms of the fully connected layer of the LSTM network are fine-tuned using the gradient descent method. The fine-tuning range is 5%-10% of the original parameters to ensure that the model can quickly adapt to production fluctuations.
[0083] Calibration frequency: Deviation detection is performed every 1.5 hours. If the triggering condition is met, calibration is performed to correct the stable output time window of molten iron.
[0084] Closed-loop simulation verification module
[0085] To ensure the feasibility of the production scheduling plan, this module simulates the supply and demand matching of molten iron under different production scheduling schemes and outputs optimization suggestions.
[0086] Simulation parameter settings
[0087] Simulation cycle: Consistent with the stable iron production time window (usually 24-48 hours);
[0088] Time unit: 30 minutes is one statistical unit;
[0089] For example, in the production of tire cord bead steel, the rules for using molten iron are as follows: Two types of molten iron are used in tire cord bead steel production: E-type and F-type. Molten iron is classified into categories A-H based on the content of common elements (C, Si, S, P, Mn) and residual elements (Cr, Ni, Cu, Mo, As, Pb, Sn, Sb, Bi, Ti), with different limits on the content of each element in each category. The rules are: only F-type molten iron is used for tire cord steel; bead steel produced earlier than tire cord steel preferentially uses E-type molten iron; if F-type molten iron has been stored for ≥3 hours, it is preferred; bead steel produced later than tire cord steel can use a mix of E and F-type molten iron; molten iron must not be stored for more than 4 hours, otherwise it must be discarded from inventory.
[0090] Simulation process
[0091] Data input: Receive the predicted or calibrated stable production time window for E / F type molten iron and the production plan for tire cord bead steel (including steel grade, production duration, and molten iron demand).
[0092] Supply and demand statistics: Based on a 30-minute time unit, the actual output of E / F type molten iron (obtained from the production scheduling and control system), the steelmaking consumption demand (calculated based on the production plan), and the molten iron inventory are calculated within that time period.
[0093] Matching degree calculation: Calculate the iron supply-demand gap ratio during the simulation period using the formula: = (Total demand - Total supply) / Total demand * 100%
[0094] Solution assessment: Ensure a certain amount of safe inventory of molten iron. If the shortage rate is <-5%, the production scheduling plan is deemed feasible, and a plan report is generated. If the shortage rate is ≥-5%, analyze the reasons for the shortage (such as insufficient supply of Class F molten iron or concentrated production time of tire bead steel), and generate optimization suggestions to coordinate with the ironmaking plant to extend the stable production time of Class F molten iron.
[0095] Prediction Results Panel
[0096] The predicted time window for stable production of E / F type molten iron is displayed in the form of a time axis, marked with green intervals;
[0097] Records the dynamic calibration history, marking calibration time points with orange dots. Hovering the mouse over the time point allows you to view the time deviation before and after calibration and the reasons for it.
[0098] The system also includes a visualization and interaction module, specifically a data monitoring dashboard, a prediction results panel, and a simulation report interface. The data monitoring dashboard displays the time-series curves of blended ore composition, blast furnace condition, and molten iron production in real time. The prediction results panel marks the stable production time window of E / F type molten iron and dynamic calibration records in the form of a time axis. The simulation report interface outputs a molten iron supply and demand matching bar chart, shortage rate statistics, and production scheduling optimization suggestions.
[0099] As one implementation method, the system also includes a risk management module, which includes:
[0100] V1: Composition monitoring unit, collects sinter composition every 4 hours and molten iron composition every 1 hour, triggering a Feishu group alert when the standard is exceeded; V2: Furnace temperature monitoring unit, collects blast furnace temperature data in real time, and pushes furnace condition adjustment suggestions to the ironmaking dispatch terminal when the furnace temperature of a single blast furnace fluctuates for 4 consecutive hours, resulting in a molten iron compliance rate of <66.7%; V3: Data quality monitoring unit, detects and addresses data acquisition distortion or delay in ladle data, and switches to the backup data acquisition channel (from the steelmaking secondary system to the production report system) when the data missing rate is >3%.
[0101] This method is applicable to the response requirements of different steel grades to different grades of molten iron.
[0102] Example 2:
[0103] like Figure 2 As shown, this embodiment provides a time-series prediction method for stable molten iron production based on LSTM, applied to the system provided in Embodiment 1, including the following steps:
[0104] Step 1: Data Acquisition and Preprocessing. The data acquisition module obtains multi-source time-series data on ore blending, blast furnace conditions, and hot metal production, then cleans, fills, and standardizes the data. Step 2: Time-Series Prediction Model Training and Inference. The preprocessed data is input into the LSTM time-series prediction module. After processing through a multi-layer LSTM network and an Attention mechanism, it outputs an E / F type stable hot metal production time window. Step 3: Dynamic Calibration. The dynamic calibration module compares the deviation between the real-time production signal and the predicted time, fine-tunes the model parameters, and corrects the time window. Step 4: Closed-Loop Simulation Verification. The closed-loop simulation verification module simulates production scheduling plans, calculates the supply-demand matching degree, and outputs feasible solutions or optimization suggestions. Step 5: Decision Support and Visualization. The visualization interaction module displays the prediction results and simulation reports, providing planners with a basis for production scheduling decisions.
[0105] In one implementation, in step 2, the feature importance ranking of the LSTM time series prediction module is as follows: Ti content of blended ore (weight 0.25-0.3) > blast furnace blast temperature (weight 0.18-0.22) > molten iron temperature (weight 0.15-0.18) > P content of blended ore (weight 0.12-0.15) > pulverized coal injection rate (weight 0.08-0.1), and the average absolute error of the model prediction time window is ≤1.5 hours.
[0106] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.
[0107] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A time-series prediction system for stable iron production from ore blending based on LSTM, characterized in that, include: The data acquisition module is used to collect multi-source time-series data on the chemical composition of blended ore, blast furnace condition parameters, and hot metal production indicators, and to preprocess the collected data. The LSTM time series prediction module, based on preprocessed multi-source time series data, captures long-term data dependencies through a multi-layer LSTM network structure, and combines an attention mechanism to enhance feature extraction for a specified process period, outputting a stable time window for molten iron to meet standards. The dynamic calibration module acquires real-time production data from the front end through a real-time data interface, compares the deviation between the actual output of molten iron and the predicted time window, and corrects the output of the LSTM timing prediction module online. The closed-loop simulation verification module is used to receive the time window output by the LSTM timing prediction module or the dynamically calibrated time window, simulate the steel production scheduling plan, calculate the iron supply and demand matching degree, and verify the feasibility of the scheduling plan.
2. The LSTM-based time-series prediction system for stable iron production from ore blending as described in claim 1, characterized in that, The data acquisition module collects multi-source time-series data, including the chemical composition data of blended ore, which includes total iron content, silicon dioxide content, titanium content, and phosphorus content; blast furnace condition parameter data, which includes furnace top temperature, blast temperature, blast pressure, oxygen enrichment rate, and pulverized coal injection rate; and molten iron production index data, which includes molten iron temperature, molten iron titanium content, and molten iron phosphorus content. The preprocessing includes data cleaning, missing value imputation, and standardization.
3. The LSTM-based time-series prediction system for stable iron production from ore blending as described in claim 1, characterized in that, The network structure of the LSTM time-series prediction module includes an input layer, multiple LSTM hidden layers, an attention layer, and a fully connected output layer. The input layer converts the preprocessed multi-source time-series data into a fixed-dimensional feature vector. In the multi-layer LSTM hidden layer, the input gate, forget gate and output gate of each LSTM unit adopt the Sigmoid activation function, the cell state update adopts the Tanh activation function, and a Dropout layer is set between adjacent LSTM layers. The attention layer enhances the feature contribution of a specified process period by calculating the weight coefficients of the features at each time step. The fully connected output layer uses a linear activation function to output the start and end times of stable molten iron production.
4. The LSTM-based time-series prediction system for stable iron production from ore blending as described in claim 1, characterized in that, The training process of the LSTM time series prediction module includes: S1. Construct a training dataset, using complete time series sequences from historical production data as samples. Each sample contains multi-source time series data features and corresponding stable iron production time labels. S2. Initialize LSTM network parameters; S3. The Adam optimizer is used to minimize the mean square error loss function, which is defined as the mean of the squared deviations between the prediction time window and the actual time window. S4. Prevent overfitting of the model by using early stopping. Stop training when the exit condition is met and save the optimal model parameters.
5. The LSTM-based time-series prediction system for stable iron production from ore blending as described in claim 1, characterized in that, The calibration logic of the dynamic calibration module includes: T1. Real-time acquisition of signals indicating completion of material replacement in the mixing bin, sinter composition compliance, and blast furnace temperature fluctuations; T2. Calculate the deviation between the actual signal trigger time and the LSTM prediction time. If the deviation is greater than the set value, start the calibration mechanism. T3. Based on the deviation value, the parameters of the fully connected layer of the LSTM network are fine-tuned using the gradient descent method to correct the stable iron production time window of the subsequent output.
6. The LSTM-based time-series prediction system for stable iron production from ore blending as described in claim 1, characterized in that, The produced molten iron includes Class I molten iron and Class II molten iron. The simulation rules of the closed-loop simulation verification module include: U1, Receive the time window output by the LSTM timing prediction module or the dynamically calibrated time window, and set the simulated start time for steel production scheduling; U2. Using a set duration as the time unit, calculate the actual output of molten iron, the steelmaking consumption demand, and the molten iron inventory within that time unit. U3. Establish priority rules for the use of Class I and Class II molten iron for different steel grades: U4. Calculate the iron supply and demand gap rate during the simulation cycle, determine whether the production scheduling plan is feasible, and output optimization suggestions.
7. The LSTM-based time-series prediction system for stable iron production from ore blending as described in claim 6, characterized in that, The system also includes a visualization and interaction module, including a data monitoring dashboard, a prediction results panel, and a simulation report interface; The data monitoring dashboard displays time-series curves of chemical composition of blended ore, blast furnace condition parameters, and hot metal production indicators in real time; the prediction results panel marks the stable hot metal production time window and dynamic calibration records in the form of a time axis; the simulation report interface outputs a hot metal supply and demand matching bar chart, shortage rate statistics, and production scheduling optimization suggestions.
8. The LSTM-based time-series prediction system for stable iron production from ore blending as described in claim 1, characterized in that, The system also includes a risk management module, which comprises: The composition monitoring unit is used to collect the composition of primary sinter and molten iron, and triggers an alarm when the composition exceeds the standard. The furnace temperature monitoring unit is used to collect blast furnace temperature data in real time. When the furnace temperature fluctuation exceeds the set condition, it pushes furnace condition adjustment suggestions to the ironmaking dispatch terminal. The data quality monitoring unit detects and addresses data acquisition distortion or delays in the water tank. When the data loss rate exceeds a set value, it switches to a backup data acquisition channel.
9. A time-series prediction method for stable iron production from ore blending based on LSTM, characterized in that, include: Step 1: Collect multi-source time-series data on chemical composition of blended ore, blast furnace condition parameters, and hot metal production indicators, and perform preprocessing. Step 2: Process the preprocessed data through a multi-layer LSTM network and attention mechanism to output the stable production time window of molten iron with ore blending; Step 3: Compare the deviation between the real-time production signal and the predicted time, fine-tune the model parameters, and correct the time window; Step 4: Based on the obtained time window closed-loop simulation verification module, simulate the production scheduling plan, calculate the supply and demand matching degree, and output feasible solutions or optimization suggestions.
10. The LSTM-based method for predicting the stable output time series of molten iron from ore blending according to claim 9, characterized in that, In step 2, the importance of features processed by the multilayer LSTM network and attention mechanism is ranked as follows: titanium content in the blended ore > blast furnace blast temperature > molten iron temperature > phosphorus content in the blended ore > pulverized coal injection rate.