Converter gas prediction and oxygen collaborative optimization method and system based on trend perception

Abstract

This invention discloses a method and system for predicting converter gas and coordinating oxygen optimization based on trend perception, belonging to the field of intelligent control and energy optimization technology. The method for predicting converter gas and coordinating oxygen optimization includes: data acquisition and preprocessing; constructing a historical trend sample set; performing real-time single-time prediction of gas based on similarity matching; multi-step rolling prediction; and inputting the obtained multi-step prediction result sequence into an oxygen dynamic control model, thereby enabling real-time control of the oxygen supply. This invention is applicable to the energy management system of converter steelmaking processes, achieving dynamic balance between oxygen production and consumption, minimizing emissions, improving energy utilization, and reducing external purchase costs. It features accurate prediction, fast response, and strong adaptability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control and energy optimization technology, and more specifically, relates to a method and system for predicting converter gas and coordinating oxygen optimization based on trend perception. Background Technology

[0002] The steel industry is a typical energy-intensive process industry, and the efficient recovery and balanced utilization of secondary energy is a core element in achieving green manufacturing and cost reduction. In the converter steelmaking process, high-purity oxygen is blown into the molten iron to oxidize and remove elements such as carbon, silicon, and phosphorus. This process generates a large amount of high-temperature converter gas (LDG), rich in carbon monoxide (CO), which has a high calorific value and is a secondary energy source with significant recovery value. Meanwhile, oxygen itself, as a key process medium, constitutes a significant portion of the energy costs for steel companies. Therefore, achieving efficient recovery of converter gas and precise control of oxygen supply has significant economic and environmental benefits for steel companies.

[0003] However, the converter gas system and oxygen system are complex dynamic systems with strong coupling and large time lag, and their stable and optimized operation has long been a technical challenge for the steel industry in my country and even the world. Existing control technologies mainly face the following difficulties: I. At the prediction level: Prediction methods based on traditional models are severely unsuitable for converter operating conditions; Failure of the mechanism model: The complex physicochemical reactions in the converter are difficult to describe online using precise mathematical models (such as equations based on mass and energy balance), and the model parameters drift with changes in raw materials, furnace conditions, and furnace age, resulting in poor accuracy and high maintenance costs for mechanism-based gas production prediction models in practical applications.

[0004] Limitations of traditional time series models: Although time series forecasting methods (such as the autoregressive integral moving average model, ARIMA) perform well in scenarios with obvious trends and periods, the converter gas generation process exhibits strong intermittency, non-stationarity and nonlinearity.

[0005] Intermittent nature: Gas recovery only occurs during the time window when the composition requirements are met in the later stages of blowing, with each cycle lasting approximately 15-20 minutes; the output is zero at other times. This "zeroing out" characteristic makes the data sequence discontinuous, violating the fundamental assumption of data stationarity in traditional time series models.

[0006] Nonlinearity: The amount of gas generated is affected by dozens of factors, including the composition of molten iron, oxygen blowing intensity, slag formation, and operating habits. There are complex interactions between these factors, resulting in a high degree of nonlinearity. Simple linear models (such as ARIMA) cannot capture its inherent laws.

[0007] Randomness: Frequent random events such as production rhythm adjustments, equipment failures, and process anomalies result in a large number of atypical "spurs" and "mutations" in the data, further increasing the difficulty of prediction.

[0008] Consequences: Because it is impossible to make accurate and forward-looking judgments on future gas production and consumption trends, the energy dispatch center can only make "post-event" and passive adjustments based on the current gas storage level and pipeline pressure, which is the direct cause of gas release or insufficient supply.

[0009] Second, at the system coordination level: the gas and oxygen systems have long been in an "information silo" state, lacking coordinated optimization; Currently, in most steel companies, gas dispatching and oxygen dispatching are managed by different work teams or departments. The goal of gas dispatchers is to stabilize pipeline pressure and reduce emissions; the goal of oxygen dispatchers is to ensure the safety of oxygen generators and meet process requirements. There is a lack of a unified information platform and collaborative optimization mechanism between the two.

[0010] This fragmented control model has resulted in enormous energy waste. For example, when multiple converters are about to finish blowing and the gas production drops sharply, the oxygen system continues to operate at high load because it fails to detect this in advance, leading to excess oxygen release; or, when there is a gas surplus, the best opportunity to generate profits is missed because the system fails to link the surplus with the power generation system based on electricity prices.

[0011] Third, at the level of control mode: it relies on human experience, has a slow response, and poor stability.

[0012] Currently, many steel companies in China still rely primarily on dispatchers' experience to determine the gas balance and issue control commands. This manual approach is limited by individual judgment, work conditions, and experience levels, resulting in slow response times; typically, a response is only possible several minutes or even longer after changes in operating conditions occur.

[0013] Regulation lag: As a large-volume buffer system, the gas pipeline network experiences pressure changes that lag behind changes in production and consumption. By the time dispatchers observe pressure anomalies, an imbalance has already occurred. This "seeing before acting" approach is perpetually outdated by changes in actual operating conditions.

[0014] Large pressure fluctuations: The "pulse-like" recovery of converter gas and fluctuations in user consumption cause the pipeline pressure to fluctuate significantly. Manual control is difficult to achieve precise and timely compensation, resulting in unstable inlet pressure for key users (such as continuous heating furnaces), affecting product quality and even causing safety accidents.

[0015] In summary, existing technologies have significant shortcomings in three dimensions—prediction accuracy, system coordination, and response speed—when dealing with the dynamic balance problem of converter gas and oxygen systems. Summary of the Invention

[0016] The first objective of this invention is to provide a trend-aware converter gas prediction method, thereby solving the technical problem that existing prediction methods based on traditional models are not applicable to converter operating conditions and have relatively poor prediction accuracy. The second objective of this invention is to provide a method for the coordinated optimization of converter gas and oxygen, thereby solving the technical problems of the complete isolation of converter gas and oxygen systems in the prior art, which leads to significant energy waste and lag in response. This invention can deeply integrate the information of the two systems to achieve accurate forward prediction and intelligent coordinated optimization. This invention also provides a converter gas prediction and oxygen co-optimization system; The present invention also provides a computer-storable medium storing a computer program thereon, which, when executed by a processor, can implement the converter gas prediction method and the converter gas and oxygen co-optimization method described above.

[0017] To achieve the above objectives, the technical solution provided by the present invention is as follows: The first aspect of this invention provides a trend-aware converter gas prediction method, comprising the following steps: Data acquisition and preprocessing: Collect coal gas data, oxygen blowing data, and process parameters related to coal gas production, and preprocess the data to obtain a standardized historical dataset; The gas data includes converter gas pipeline pressure, converter gas instantaneous flow rate, and cumulative recovery amount; the oxygen blowing data includes converter oxygen blowing instantaneous / cumulative flow rate and oxygen main pipeline pressure; the process parameters related to gas generation include molten iron feed rate, scrap steel quantity, molten iron temperature, blowing time, and smelting cycle status indicator; the preprocessing includes data cleaning, data filtering, and normalization. Data cleaning includes missing value imputation and outlier identification and processing. Data filtering uses median average filtering to smooth noise.

[0018] Dynamic extraction of the comprehensive physical lag constant: Based on the historical dataset, calculate the maximum cross-correlation function between the total oxygen blowing flow sequence and the total gas recovery flow sequence of multiple furnaces, and determine the delay step size corresponding to the extreme point of the maximum cross-correlation function as the comprehensive physical lag constant of the current converter cluster. ; Furthermore, the formula for the maximum cross-correlation function between the total oxygen blowing flow rate sequence and the total gas recovery flow rate sequence of the multiple furnaces is as follows:

[0019] in, For the first The total oxygen blowing flow rate of all converters in the plant is monitored at all times. For delay Total gas recovery flow rate after time step The total length of the sample. This is the overall physical hysteresis constant.

[0020] Adaptive Prediction Window Setting: Set the Length of Historical Trend Comparison and future predicted length and make the future prediction length The corresponding prediction duration is greater than or equal to the composite physical lag constant. .

[0021] Constructing a historical trend sample set: The standardized historical dataset is divided using a sliding window technique, and each window index is... , cut The data in the intervals form a historical comparison sequence, and the data is extracted. The data within the interval constitutes the future prediction sequence. All historical comparison sequences are associated with the future prediction sequence to form a historical trend sample set. Furthermore, the data in the historical comparison sequence includes gas data, oxygen blowing data, and process parameters at various time points; the future prediction sequence only includes converter gas recovery data; and the historical trend sample set is updated offline periodically.

[0022] Real-time single-shot prediction based on similarity matching: truncate the latest data from the end of the current historical dataset to a length of... The data is used to construct a real-time query sequence; the similarity distance between this real-time query sequence and all historical comparison sequences in the historical trend sample set is calculated, and the top... The future prediction sequences corresponding to the most similar historical comparison sequences are fused and calculated to output a single prediction result. Furthermore, the similarity distance is calculated using Manhattan distance or Euclidean distance; the fusion calculation employs a weighted fusion algorithm based on similarity distance, in which historical sequences with smaller similarity distances are assigned greater weights.

[0023] Multi-step rolling prediction: The first predicted value from a single prediction is added as a "pseudo-actual value" to the end of the historical dataset, while the oldest data point at the beginning of the dataset is removed to update the dataset; the real-time single prediction is repeated using the updated dataset until a value of length is obtained. The multi-step prediction result sequence.

[0024] A second aspect of the present invention also provides a method for coordinated optimization of converter gas and oxygen, which determines the oxygen supply adjustment amount based on a multi-step prediction result sequence obtained from the prediction, specifically including the following steps: Obtain the multi-step prediction result sequence obtained by the method described in the first aspect of the present invention, and the comprehensive physical hysteresis constant of the current converter cluster. ; Cross-medium oxygen consumption feedforward mapping: based on the aforementioned integrated physical hysteresis constant The predicted gas production value at future moments within the forecast period is reverse-mapped to the predicted dynamic oxygen consumption of the converter at that future moment. The core logic of the inverse mapping is: the dynamic oxygen consumption of the converter in the future, and the lag... There is a strong positive correlation between the predicted gas production value after a certain time. By fitting historical data, a linear mapping function between the two can be obtained, realizing the reverse derivation from the predicted gas production value to the predicted oxygen consumption value.

[0025] Construct an oxygen dynamic control model: Using the oxygen supply regulation amount as the control variable, construct an oxygen dynamic control model that includes dynamic collaborative constraints on oxygen pipeline pressure; the dynamic collaborative constraints on oxygen pipeline pressure are: at any future time within the prediction period, based on the predicted value of the converter's dynamic oxygen consumption and the control variable, calculate the total oxygen pipeline pressure and restrict it to meet the safety boundary. Furthermore, the mathematical model for the dynamic collaborative constraint of the oxygen pipeline network pressure is as follows:

[0026] In the formula, , These are the minimum and maximum safe pressures allowed by the oxygen pipeline network, respectively. This represents the actual pressure of the oxygen pipeline network at the current moment. The volume constant of the oxygen pipeline network; The dynamic supply flow rate of the oxygen generator at time i in the future is the model control variable. This is the predicted dynamic oxygen consumption of the converter at time i in the future. For time step.

[0027] Furthermore, the specific execution logic of the dynamic collaborative constraint of the oxygen pipeline pressure is as follows: when it is determined from the predicted value of gas production that the dynamic oxygen consumption of the converter will drop sharply at a certain time in the future, the dynamic supply flow of the oxygen generator in the control variable is forcibly reduced in advance to prevent the total oxygen pipeline pressure at that future time from exceeding the upper limit of the safety boundary.

[0028] Model Solving and Control: Under the conditions of satisfying the constraints of gas user priority, gas storage location, and gas pipeline pressure, the oxygen dynamic control model is solved with the objectives of minimizing gas emissions and maximizing gas utilization benefits, and the oxygen supply adjustment amount at each future time point is output.

[0029] A third aspect of the present invention also provides a converter gas prediction and oxygen co-optimization system, comprising: The data acquisition and preprocessing module is used to collect coal gas data, oxygen blowing data, and process parameters, and to preprocess them to obtain a standardized historical dataset. The hysteresis constant extraction module is used to calculate the maximum cross-correlation function between the total oxygen blowing flow rate and the total gas recovery flow rate based on the historical dataset, and dynamically determine the comprehensive physical hysteresis constant of the current converter cluster. ; The historical trend sample set management module is used to set the length of future predictions. Make it greater than or equal to And a historical trend sample set is constructed using sliding window technology; The real-time single prediction module is used to perform real-time single prediction based on similarity matching and output the single prediction result; the multi-step rolling prediction module is used to obtain a multi-step prediction result sequence of length M by iteratively updating the dataset with the predicted values ​​as pseudo-actual values.

[0030] Furthermore, the system also includes an oxygen dynamic control module, which comprises: Feedforward mapping unit, used for based on the comprehensive physical hysteresis constant The predicted value of gas production at future moments is mapped inversely to the predicted value of dynamic oxygen consumption of the converter. The collaborative constraint solving unit is used to introduce dynamic collaborative constraints on oxygen pipeline pressure based on the predicted oxygen consumption value into the oxygen dynamic regulation model, and solve for the output oxygen supply regulation amount under the condition of satisfying multi-objective optimization.

[0031] A fourth aspect of the present invention also provides a computer-storable medium having a computer program stored thereon, which, when executed by a processor, can implement the converter gas prediction method as described in the first aspect of the present invention, or the oxygen co-optimization method as described in the second aspect of the present invention.

[0032] The fifth aspect of the present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it can implement the converter gas prediction method as described in the first aspect of the present invention, or the oxygen co-optimization method as described in the second aspect of the present invention.

[0033] Compared with the prior art, the technical solution provided by this invention has the following advantages: (1) High prediction accuracy and strong adaptability: The present invention adopts a converter gas prediction method based on historical trend similarity, which does not require a complex mechanism model. It can effectively capture the intermittent and nonlinear characteristics of converter smelting, and has stronger adaptability to complex working conditions. The prediction accuracy is significantly higher than that of traditional time series models.

[0034] (2) Overcoming the defects of fixed mechanistic model parameters and achieving strong adaptability and generalized deployment capability for equipment of different specifications: In the actual production of converter steelmaking, the physical lag constant of 'oxygen blowing-gas production' is not fixed, but is strictly constrained by underlying physical conditions such as converter tonnage specifications (e.g., differences in molten pool depth), length of dust removal pipeline layout, oxygen supply intensity, and mechanical delay of field instrument valves. When traditional mechanistic prediction models are deployed across equipment and plant areas, a large number of physical parameters must be remeasured and cumbersomely calibrated, resulting in extremely high engineering maintenance costs.

[0035] This invention employs a data-driven historical trend similarity matching mechanism. By setting a sliding window that includes the complete physical delay period (ensuring that the prediction step size M covers or is greater than the hysteresis constant), the model can implicitly and accurately absorb and fit the physical delay characteristics specific to the current converter when comparing historical waveforms. This means that the system does not need prior physical parameters such as specific pipe lengths or furnace volumes; it only needs to rely on the historical operating data of the equipment to automatically adapt to its unique dynamic hysteresis patterns. This mechanism completely decouples the strong binding relationship between the prediction algorithm and the underlying hardware physical specifications, achieving efficient "plug-and-play" across devices and specifications, and significantly improving the system's generalization ability and engineering application value.

[0036] (3) Achieving true collaborative optimization: This invention constructs a coal gas-oxygen joint trend sample and considers the goals and constraints of the two systems in a unified manner in the optimization model, breaking the information silos between the systems and enabling precise matching of energy and maximization of benefits from a global perspective.

[0037] (4) Rapid response and advanced regulation: The present invention adopts multi-step rolling prediction, which can predict the gas trend in the future period of time, thereby providing valuable advance for the regulation of oxygen system and buffer users, changing passive response to active advanced regulation, effectively suppressing pipeline pressure fluctuations, and solving the technical problems of large energy waste and slow response in the existing technology.

[0038] (5) This invention introduces oxygen blowing data as a leading feature into the historical comparison sequence and utilizes the strong physicochemical correlation between “oxygen blowing-decarbonization-gas production” in the converter smelting process. This enables the model to detect the gas production trend in advance by identifying the oxygen blowing mode before the actual change in the amount of gas produced (with a lag time of about several minutes), thereby completely solving the problem of lag in traditional feedback control.

[0039] (6) Self-learning and continuous improvement: The system design includes a feedback loop, which can continuously update the historical trend sample set using actual operating data, so that the model can continuously learn new production rules, has the ability to self-evolve, and maintain excellent performance in the long term. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall architecture of the converter gas and oxygen co-optimization system provided in an embodiment of the present invention.

[0041] Figure 2 This is a flowchart illustrating the process of constructing a historical trend sample set in an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of the single-step prediction principle based on similarity matching in an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram of the multi-step rolling prediction process in an embodiment of the present invention.

[0044] Figure 5 This is a logic block diagram of the oxygen dynamic regulation model in this embodiment of the invention.

[0045] Figure 6 This is a complete flowchart of the converter gas and oxygen co-optimization method provided in the embodiments of the present invention. Detailed Implementation

[0046] The embodiments of this disclosure will now be described in conjunction with specific examples. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of this disclosure.

[0047] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0048] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0049] This invention provides a trend-aware converter gas prediction method and an oxygen collaborative optimization method. Its core lies in achieving intelligent linkage between the converter gas system and the oxygen system through a closed-loop process of "data preprocessing → dynamic lag extraction → adaptive window setting → sample library construction → similarity prediction → rolling prediction → collaborative optimization." Figure 1 , Figure 6 As shown, the method specifically includes the following steps: S1. Multi-source heterogeneous data acquisition and fusion preprocessing; Through a sensor network (such as flow meters, pressure transmitters, and component analyzers) deployed at key nodes such as converters, gas holders, and oxygen pipelines, and through interfaces with the enterprise's Manufacturing Execution System (MES), four-dimensional data is collected in real time. "Oxygen blowing data" is explicitly included in the data collection process. The complete data collection dimensions are as follows: Gas data: including instantaneous / cumulative flow rate of converter gas, gas holder level, and gas calorific value.

[0050] Oxygen data: including instantaneous / cumulative oxygen flow rate and oxygen pipeline pressure.

[0051] Process data includes flue gas temperature, steam flow rate, flue gas CO content, oxygen lance flow rate, and status indicators for each smelting stage (such as blowing, settling, and tapping).

[0052] Planned data: Dynamic furnace output plans, steel grade information, and maintenance plans from the MES (Manufacturing Execution System).

[0053] The high-frequency time series data mentioned above is obtained from the underlying PLC and the first-level instrument system through the OPC interface, with the sampling frequency set to 1 second.

[0054] The collected raw data can undergo data alignment (unifying to the same timestamp), data cleaning (such as using the KNN algorithm to fill missing values ​​and using box plots to identify and process outliers), data filtering (such as using median average filtering to smooth noise), and normalization to ultimately form a standardized, high-quality historical dataset.

[0055] S2. Dynamically extract the comprehensive physical hysteresis constant ; This step is the core preliminary step of the prediction method. Based on the standardized historical dataset obtained by S1, the actual physical lag of the current converter cluster is dynamically extracted in a data-driven manner, which completely solves the industry pain point of the disconnect between traditional fixed experience delay and actual working conditions.

[0056] Specifically, calculate the maximum cross-correlation function between the total oxygen blowing flow rate sequence and the total gas recovery flow rate sequence of multiple furnaces. The delay step size corresponding to the extreme point of the maximum cross-correlation function is determined as the comprehensive physical lag constant of the current converter cluster. That is, by searching within the set maximum search delay step, it is possible to find the solution that makes the search more efficient and efficient. Reaching the extreme value (maximum value) This allows for precise identification of the complex fluid lag under multi-furnace cross-blowing from a purely mathematical perspective. The lag step corresponding to this extreme point is the comprehensive physical lag constant of the current converter cluster.

[0057] The formula for the maximum cross-correlation function between the total oxygen blowing flow rate sequence and the total gas recovery flow rate sequence of the multi-furnace system is:

[0058] in, For the first The total oxygen blowing flow rate of all converters in the plant is monitored at all times. For delay Total gas recovery flow rate after time step The total length of the sample. This is the overall physical hysteresis constant.

[0059] This invention does not employ a fixed empirical delay, but rather dynamically extracts the actual physical lag through a data-driven approach. The physical significance lies in its ability to precisely pinpoint the complex fluid latency under multi-furnace cross-blowing conditions from a purely mathematical perspective. In this embodiment, by calculating historical data from the plant's three 150-ton converters, the comprehensive physical lag constant of the current converter cluster is ultimately obtained. =2.5 minutes.

[0060] S3. Adaptively set the prediction window; This step is based on the comprehensive physical hysteresis constant extracted by S2. Adaptively set the core window parameters of the prediction model, specifically including: S31. Set the length of historical trend comparison. Based on the average cycle time of converter smelting (approximately 30-40 minutes), this embodiment sets... This corresponds to a review of the time sequence characteristics over the past 60 minutes, fully covering the core change patterns of a complete smelting cycle in the converter; S32, Set the future forecast length Strictly follow The mathematical boundaries are defined to ensure that the prediction step size fully covers the comprehensive physical lag period. This embodiment sets... This corresponds to predicting the trend over the next 15 minutes, satisfying... minute Adaptive setting requirements in minutes.

[0061] This adaptive setting mechanism ensures that the prediction window is perfectly matched with the actual physical lag characteristics of the current converter cluster, completely solving the problem of prediction accuracy decay caused by the disconnect between the traditional fixed window and the actual working conditions.

[0062] S4. Construction of historical trend sample set; This step, based on the window parameters set in S3, constructs a standard sample library for similarity matching using the sliding window technique, combined with... Figure 2 As shown, it specifically includes: S41. Define the standardized historical dataset after preprocessing in S1; S42, Sliding window segmentation and sequence construction; Constructing a historical comparison sequence Starting from index i, extract a continuous data segment of length L to form the input feature matrix, i.e., the historical comparison sequence; Constructing future prediction sequences From the index Initially, a continuous data segment of length M is extracted, and the target variable (converter gas recovery amount) is extracted to form the target vector, i.e., the future prediction sequence; S43. Sample set generation: This involves generating historical comparison sequences. With future prediction sequence Establish mapping relationships to form sample pairs Iterate through all indices i, removing those whose length is insufficient due to device downtime or missing data segments. The invalid samples are used to form a historical trend sample set containing K valid samples. ; This sample set is updated daily offline to accommodate distribution drift caused by furnace aging and process adjustments.

[0063] S5. Real-time single-shot prediction based on similarity matching; Combination Figure 3 As shown, at the current time t when regulation is required, the following real-time single-prediction steps are performed: S51. Construct a real-time query sequence; extract the latest data of length L from the end of the current historical dataset to construct a real-time query sequence. ; S52. Similarity distance calculation; using Manhattan distance or Euclidean distance formulas, calculate... Comparison sequence with each historical trend sample set The similarity distance; S53, Generation of weighted fusion prediction results.

[0064] S6, Multi-step rolling forecast; To improve the long-term accuracy of multi-step prediction, this embodiment adopts an iterative rolling prediction strategy, combined with... Figure 4 As shown, it specifically includes: S61. Initialize an empty multi-step prediction result sequence. ; S62. Execute the single prediction result sequence obtained in S5, extract its first predicted value, and append it to the result sequence. middle; S63. The predicted value is appended to the end of the historical dataset as a "pseudo-actual value", while the oldest data in the dataset is removed to achieve the rolling update of the data window. S64. Based on the updated dataset, repeat steps S5 (real-time single prediction) to S63 (data window rolling update); iterate in this way until the result sequence is obtained. The sequence length in the sequence reaches the future prediction length M.

[0065] Finally, the result sequence The data stored in the middle is the final multi-step prediction result for the next M time points, obtained through M iterations.

[0066] S7, oxygen synergistic optimization regulation; Combination Figure 5 As shown, the final multi-step prediction results are input into the oxygen dynamic regulation model, which is a multi-objective optimization model: (1) Objective function: Objective 1: Minimize the amount of coal gas emitted; Objective 2: Maximize the efficiency of coal gas utilization, that is, prioritize the use of surplus coal gas for power generation during periods of high electricity prices.

[0067] (2) Constraints: User priority: Users are divided into three categories. Category 1 users (such as hot blast stoves) must be guaranteed supply; Category 2 users (such as steel rolling heating furnaces) can be adjusted through negotiation; Category 3 users (such as self-owned power plants and gas holders) are the main adjustment units. Equipment safety: Gas holder positions, gas pipeline pressure, oxygen pipeline pressure, etc., must be within safe upper and lower limits; Process requirements: "Zero rejection" of converter gas; Solution Algorithm: The model is solved using optimization algorithms such as the simplex method, and the final output is the oxygen supply adjustment suggestions, power generation load adjustment suggestions, and gas holder position control suggestions for the next M time points.

[0068] Specifically, the process of coordinated optimization and regulation of oxygen in S7 includes the following steps: Step 1: Cross-medium oxygen consumption feedforward mapping; The system acquires a multi-step prediction result sequence and a comprehensive physical lag constant. Subsequently, due to the strict physical stoichiometric relationship of the carbon-oxygen reaction, the system... Using time anchors, future moments within the prediction period will be... The predicted value of gas production is used to reverse-map and deduce the predicted value of converter dynamic oxygen consumption at that future moment.

[0069] Step 2: Construct a dynamic oxygen regulation model; Using the oxygen supply regulation (such as the oxygen generator guide vane opening) as the control variable, a dynamic coordination constraint on oxygen pipeline pressure is introduced into the model. The dynamic coordination logic is as follows: when it is determined from the above feedforward mapping that the converter's dynamic oxygen consumption will drop sharply due to the shutdown of blowing at a certain time in the future, in order to prevent the total oxygen pipeline pressure from exceeding the upper limit of the safety boundary, the model forces the dynamic supply flow of the oxygen generator to be adjusted down in advance at the current time.

[0070] Step 3: Model Solving and Control; Under the premise of satisfying multiple physical constraints such as user priority, gas holder location and pipeline pressure, the objective functions are to minimize the amount of gas released (objective one) and maximize the utilization benefits (objective two). Solvers such as MILP are called to output the optimal oxygen supply regulation at each future time.

[0071] The oxygen blowing mode within a smelting cycle directly determines the gas production curve within that cycle. However, in current control strategies, oxygen supply planning is not used as a key input feature for gas prediction; conversely, gas prediction results are not used to guide the proactive adjustment of the oxygen system (such as adjusting the oxygen generator load or liquid oxygen vaporization rate in advance), resulting in significant energy waste and technical problems of response lag. Based on the above, this invention constructs a gas-oxygen joint trend sample, uses gas prediction results to adjust parameters such as oxygen supply in real time, and considers the objectives and constraints of both systems uniformly in the optimization model, breaking down information silos between systems. This enables precise energy matching and efficiency maximization from a global perspective, allowing the model to detect gas production trends in advance by identifying the oxygen blowing mode before actual changes in gas production (with a lag of approximately several minutes), thus effectively solving the problem of lag in traditional feedback control.

[0072] In summary, this invention constructs a joint trend sample set of coal gas and oxygen, combining similarity matching and multi-step rolling prediction techniques to dynamically optimize oxygen supply strategies, thereby maximizing coal gas recovery efficiency and minimizing emissions. This invention is applicable to energy management systems in converter steelmaking processes, enabling dynamic balance between oxygen production and consumption, minimizing emissions, improving energy utilization, and reducing external purchase costs. It features accurate prediction, fast response, and strong adaptability.

[0073] Accordingly, the present invention also provides a converter gas prediction and oxygen co-optimization system, the system comprising: Data acquisition and interface module: responsible for acquiring real-time data from the underlying control system and MES; The data acquisition and preprocessing module is used to collect coal gas data, oxygen blowing data, and process parameters, and to preprocess them to obtain a standardized historical dataset. The hysteresis constant extraction module is used to calculate the maximum cross-correlation function between the total oxygen blowing flow rate and the total gas recovery flow rate based on the historical dataset, and dynamically determine the comprehensive physical hysteresis constant of the current converter cluster. ; Historical trend sample set management module: used to set the length of future predictions. and make it greater than or equal to And a historical trend sample set is constructed using sliding window technology; The real-time single prediction module is used to perform real-time single predictions based on similarity matching and output the single prediction results. The multi-step rolling prediction module is used to obtain a multi-step prediction result sequence of length M by iteratively updating the dataset with predicted values ​​as pseudo-actual values. Furthermore, the system also includes an oxygen dynamic regulation module, which includes: The feedforward mapping unit is used to back-map the predicted value of gas production at future times to the predicted value of dynamic oxygen consumption of the converter based on the comprehensive physical lag constant τ. The collaborative constraint solving unit is used to introduce dynamic collaborative constraints on oxygen pipeline pressure based on the predicted oxygen consumption value into the oxygen dynamic regulation model, and solve for the output oxygen supply regulation amount under the condition of satisfying multi-objective optimization.

[0074] Example 1 This embodiment uses a steelmaking plant of a large steel enterprise as the application scenario. The plant has three 150-ton converters, and the gas recovery system is equipped with electrostatic precipitators and gas holders. Oxygen is supplied by the plant's oxygen generator unit.

[0075] like Figure 1 As shown in the figure, the data-driven method for coordinated optimization of converter gas and oxygen according to an embodiment of the present invention specifically includes the following steps: Step 1: Data Acquisition and Preprocessing; High-frequency time-series data is acquired from the underlying PLC and primary instrumentation system via the OPC interface, with a sampling frequency set to 1 second. The acquired data includes: converter gas pipeline pressure, converter gas instantaneous flow rate, oxygen main pipeline pressure, converter oxygen blowing flow rate, etc. Simultaneously, production process parameters are acquired from the MES system, including molten iron feed rate, scrap steel quantity, molten iron temperature, blowing time, and current smelting cycle status indicators (such as: charging, blowing, tapping, slag splashing, etc.).

[0076] The collected raw data is cleaned and standardized to obtain a standardized historical dataset, which includes: 1.1 Missing value imputation: To address data loss caused by interrupted signal transmission or instrument malfunction, the system first determines the time span of the missing data. If the missing time is less than a preset threshold (e.g., 5 sampling periods), a time-weighted K-nearest neighbor (KNN) interpolation algorithm is used to fill the gap. The calculation formula is as follows: (1); in, is the fill value at time t; K is the number of neighboring valid data points selected (in this embodiment, K=2, that is, 2 points before and after are selected); x (t±i) These are valid measured values ​​for the nearest time. w i The weighting coefficient is inversely proportional to the time interval and is defined as follows: .

[0077] 1.2 Outlier Identification and Handling: An improved boxplot method was used to detect anomalies in the imputed data. First, the lower quartiles of the data within the sliding window were calculated. Q 1 (25th percentile) and upper quartile Q 3 (75th percentile), calculate interquartile range The threshold for identifying outliers is calculated using the following formula: (2); (3); in, V upper This indicates the upper bound for anomaly detection. V lower This indicates the lower bound for anomaly detection.

[0078] If the data is detected at a certain moment x t > V upper or x t < V lower If the value is not found, it is considered an outlier and corrected using the valid value from the previous time step.

[0079] 1.3 Data Smoothing Filtering: To address the high-frequency random noise present in the converter gas flow and pressure data, a median average filtering algorithm to prevent pulse interference is employed. The calculation logic is as follows: (4); in, x’ t The smoothed value after filtering at time t; S t For the current time, there is a set of sampled sequences of length N. x t , x t-1 ,…, x t-N+1 (In this embodiment, N=10) x max and x min Sets S t The algorithm identifies the maximum and minimum values ​​in the data. It removes extreme values ​​while preserving the true trend of the data.

[0080] Step 2: Constructing a historical trend sample set; like Figure 2 As shown, this step aims to construct a standard sample library (or sample set) for similarity matching, specifically including:

[0081] 2.1 Parameter initialization and definition; The data after all the preprocessing steps, including missing value imputation, outlier handling, and filtering smoothing, are integrated to define the standardized historical dataset as follows: For the sake of simplicity, the preprocessed multidimensional feature vectors will be uniformly referred to as follows: Where N is the total data duration, It is a multidimensional feature vector at time t (containing coal gas, oxygen and process parameters).

[0082] Based on the average cycle of converter smelting (approximately 30-40 minutes) and the lag characteristics of oxygen regulation, the historical trend comparison length is set as L (in this embodiment, L=60, representing the time series characteristics of the past 60 minutes), and the future prediction length is set as M (in this embodiment, M=15, representing the predicted trend of change in the next 15 minutes). When setting the historical trend comparison length L and the future prediction length M, the system fully considers the physical lag constant of the converter steelmaking process of 'oxygen blowing first, gas production later'. In actual industrial settings, this constant is not a fixed value, but is dynamically determined by four major physical factors: (1) furnace specifications and molten pool depth; (2) fluid transmission delay caused by the physical spatial layout of the pipeline network and the length of the pipeline; (3) differences in oxygen supply intensity and smelting stage for different steel grades; and (4) delay in the mechanical action of the gas analyzer sampling and the three-way switching valve.

[0083] More complexly, when multiple converters operate simultaneously throughout the plant, nonlinear hydrodynamic collisions and air resistance effects occur at the intersection of the main pipeline network, forming a "comprehensive physical lag constant" that is difficult to solve precisely using traditional mathematical equations. This invention ensures the prediction step size. The data window design enables the system to implicitly and accurately absorb and fit the complex physical delay characteristics specific to the current converter cluster when performing waveform similarity matching. This mechanism, which explicitly incorporates the lag of the process mechanism into the sliding window scale design, constitutes the physical basis of the feedforward advance control of this invention. The standardized historical dataset is defined as D={ x 1, x 2,…, x N}, where N is the total data duration. x t It is a multidimensional feature vector at time t (containing coal gas, oxygen and process parameters).

[0084] This invention does not employ a fixed empirical delay, but rather dynamically extracts the comprehensive physical lag constant by calculating the maximum cross-correlation function (CCF) between the total oxygen blowing flow rate sequence and the total gas recovery flow rate sequence of multiple furnaces before constructing the historical trend sample set offline. Its mathematical model is defined as follows: (5); in, y represents the total oxygen blowing flow rate of all converters in the plant at time k; For delay Total gas recovery flow rate after time step; N is the total sample length; The maximum search delay step size is set. The physical meaning of formula (5) is that it accurately locks the complex fluid comprehensive delay under multi-furnace cross-blowing from a purely mathematical perspective. This serves as a rigorous mathematical boundary for setting the future prediction length M (i.e., ensuring...). ).

[0085] The aforementioned maximum cross-correlation function (CCF) model is not merely a mathematical calculation, but rather a digital representation of the coupling hysteresis characteristics of converter gas-solid-liquid multiphase flow. The specific correspondence between its underlying computational principle and physical meaning is as follows: (1) Physical mapping of corresponding dot product operators (dot product mechanism): core terms in the formula Mathematically, this is the dot product operation of two discrete signals. In a physical scenario, the peak of oxygen blowing (the precursor signal) is inevitably accompanied by the intensification of the carbon-oxygen reaction, leading to a peak in gas production (the response signal). The dot product operation utilizes the nonlinear mathematical property that "multiplying large numbers causes significant numerical amplification." When the assumed time shift... When the peaks in two sequences perfectly overlap on the time axis, the dot product will be exponentially amplified; conversely, if the peaks and troughs are misaligned, the product will be greatly weakened. This operation, at its underlying logic, acts as a detector of the causal coupling strength between oxygen supply and gas production within a fluid conduit.

[0086] (2) Anti-interference mechanism of global integral summation (Σ): The sequences acquired by sensors in industrial fields are often accompanied by strong electromagnetic interference and instantaneous spikes (such as single-point extreme value anomalies). The formula uses an integral summation operator. For the entire sample window The mapping results within the region are globally accumulated. This mechanism can effectively suppress the interference of local noise on the correspondence between single points, ensuring that the calculated similarity score reflects the inevitable physical law of the global trend, rather than local random errors, and greatly improves the robustness of the algorithm under harsh working conditions.

[0087] (3) Boundary of extreme value search ( Convergence of the project: operator Drive system at the maximum physical delay boundary allowed by engineering Within this process, the system automatically optimizes to find the specific translation time that maximizes the absolute value of the global coupling product score. This time solution, mined through pure data-driven analysis, represents the true comprehensive physical lag constant under the current complex pipeline network conditions of the entire plant. .

[0088] In summary, this formula cleverly transforms the complex problem of fluid dynamics mapping into an optimization problem of time-domain translation and vector dot product of a one-dimensional discrete time series, fundamentally overcoming the industry technical bottleneck that traditional mechanism models cannot accurately quantify lag time under multi-furnace cross-operation.

[0089] 2.2 Sliding window segmentation and sequence construction; The dataset D is traversed and segmented using a sliding window with a step size of 1. For any time index i (1≤i≤NL-M+1), the corresponding sample unit is constructed.

[0090] (1) Constructing a historical comparison sequence P i Starting from index i, extract a continuous data segment of length L to form the input feature matrix, represented as: (6); in,P i It includes the complete evolution trajectory of gas generation, oxygen consumption, and process status over the past L time points.

[0091] (2) Constructing future prediction sequences Q i Starting from index i+L, extract a continuous data segment of length M, and extract the target variable (converter gas recovery amount) from it to form a target vector, represented as: (7); in, y i This represents the numerical vector at time t containing only the predicted targets (gas and oxygen).

[0092] (3) Sample set generation: Generate historical comparison sequences P i With future prediction sequence Q i Establish mapping relationships to form sample pairs { P i , Q i}. Traverse all indices i, removing invalid samples whose length does not meet L+M due to equipment downtime or missing data segments, ultimately forming a historical trend sample set S={{ P 1, Q 1},…, { P K , Q K}}.

[0093] The sample set is further optimized to be updated daily in an offline manner to adapt to the distribution drift caused by furnace aging and process adjustments.

[0094] Step 3: Real-time single prediction; like Figure 3 As shown, at the current time t when regulation is required: 3.1 Constructing the current feature set: Extract the latest data of length L from the end of the current historical dataset to construct the real-time query sequence. P current .

[0095] 3.2 Similarity Calculation: Calculation P current Comparison sequence with each historical trend sample set P i The similarity distance. This embodiment uses the Manhattan distance formula: (8) in, P current Represents a real-time query sequence. P i The i-th historical comparison sequence represents the historical trend sample set. p current,j and p i,j Sequences P current and P i The data point at the j-th time position (i.e., the normalized value). Distance D Manhattan The smaller the value, the more similar the current operating conditions are to those historical operating conditions.

[0096] 3.3 Generation of prediction results: The Manhattan distance is calculated according to formula (8). D Manhattan All historical sequences are sorted in ascending order, and the top K (K=5 in this embodiment) historical comparison sequences with the smallest distance are selected as the similarity set. To improve prediction accuracy, this embodiment of the invention uses a weighted fusion algorithm based on similarity distance to calculate the predicted value. The specific calculation model is as follows: First, calculate the weight of the k-th similar historical sequence. W k The calculation formula is: (9); in, D k This represents the Manhattan distance between the k-th similar historical sequence and the real-time query sequence; ε To prevent the smoothing factor from being zero in the denominator, this embodiment takes a value of 1×10. -6 .distance D k The smaller, the higher the weight W k The larger the value, the more valuable the historical operating conditions are for future reference.

[0097] Subsequently, a weighted average is calculated based on the weights of the future prediction sequences corresponding to the K similar historical sequences to obtain the prediction value of the current time t for the m-th future time. P pred (t+m), the calculation formula is: (10); Where m represents the prediction step size index (m=1, 2, ..., M); F k (t+m) represents the actual value of the future prediction sequence associated with the kth similar historical sequence at time m; This is the weight normalization factor. The system outputs a predicted sequence containing M future time points using formula (10). .

[0098] Step 4: Multi-step collaborative optimization like Figure 4 As shown, in order to solve the problem of the decrease in accuracy of single prediction over time and to achieve closed-loop control of the system, this step adopts a strategy that combines rolling prediction with optimization solution.

[0099] 4.1 Multi-step rolling prediction mechanism; To obtain the system state at M future time points, an iterative rolling prediction strategy is adopted, and the specific process is as follows: The initial prediction sequence is empty. In the m-th iteration (1≤m≤M), the weighted fusion model calculates the single-step prediction value. Subsequently, this predicted value was... Real-time query sequence as "pseudo-observations" x (t+m-1) Construct new real-time query sequences x (t+m) The prediction model is called again. This process is repeated until a predicted sequence of gas production for the next M time points is generated. .

[0100] 4.2 Construction of a dynamic oxygen regulation model: Predicted sequence V prod The input is fed into the oxygen dynamic regulation model. This model is defined as a multi-objective constrained optimization problem, aiming to solve for the optimal sequence of control variables. (Including oxygen regulation and power generation load).

[0101] (1) Set of constraints: The system operation must meet physical safety and process boundary requirements, and the constraints are defined as follows: Gas holder position safety constraints: the position of the gas holder at any time t within the prediction period. S t Must meet: (11); in, S min , S max These are the lower and upper limits for the safe operation of the gas holder, respectively. v prod , v cons , v power These are respectively the gas production, consumption, and power generation flow rates; For time step.

[0102] Pipeline pressure balance constraints: Pressure P at critical nodes gas (t) needs to be maintained within the dynamic equilibrium range: P min ≤ P gas (t)≤ P max (12); Among them, P min and P max This represents the extreme value of pressure fluctuation allowed by the process of this production line.

[0103] Dynamic Coordination Constraints on Oxygen Pipeline Pressure: At any future time t+m within the prediction period, the total oxygen pipeline pressure... The safety boundary must be satisfied. Its collaborative constraint model is expressed as: (13); in, The volume constant of the oxygen pipeline network; The dynamic supply flow rate (control variable) for the oxygen concentrator; This refers to the dynamic oxygen consumption of the converter.

[0104] Feedforward collaborative logic: due to future oxygen consumption U consume and lag τ 综合 Predicted gas production after a certain time V prod There exists a strong positive correlation mapping. When the system predicts the future t+m+... τ 综合 When the coal gas production approaches zero, a sharp drop in oxygen consumption will be detected simultaneously. To prevent the oxygen pressure in the above equation from exceeding the upper limit and causing a release, the optimized solver will force a premature reduction in the control variables. U supply (For example, reducing the opening of the oxygen generator guide vanes), thereby realizing a feedforward-feedback collaborative closed loop at the mathematical level, which uses gas prediction to drive oxygen supply in reverse.

[0105] (2) Objective function set: Objective 1: Minimize gas emissions (14); in, V t, prod Let be the predicted gas production value for period t. V t, cons These are the predicted gas consumption values ​​for Class I and Class II users in period t. V t, power Let be the amount of gas consumed by the power plant in period t. The change in gas holder level during period t; by adjusting the power generation and consumption. V t,power and gas holder level change This causes the total emission of the system to approach zero.

[0106] Objective 2: Maximize efficiency (15); in, k For gas power generation efficiency, V t, peak and V t, off-peak These represent the surplus gas power generation during peak and off-peak periods, respectively. Price peak and Price off-peak These are peak-hour and off-peak electricity prices, respectively.

[0107] Under the premise of meeting objective one, priority should be given to peak electricity price periods ( Price peak Increase the amount of gas generated and store gas in gas holders during off-peak electricity pricing periods.

[0108] Driven by the aforementioned goal of maximizing benefits, the system essentially treats the gas holder as a "virtual energy storage battery." The optimization solver incorporates a dynamic boundary strategy based on time-of-use pricing: when the multi-step rolling prediction determines that the system is about to transition from the "flat / valley pricing period" to the "peak pricing period," the model will actively tighten the lower limit for the safe operation of the gas holder in formula (11). S min (For example, raising it to 40% of the warning level) forces the system to "store energy" during periods of low electricity prices; however, once peak electricity prices arrive, the model will automatically relax the restrictions. S min Constraints instruct power plants to operate at full capacity to generate electricity from coal gas for "energy arbitrage." This dynamic interplay between energy storage and power generation constitutes the core execution logic for optimizing the overall economy of this system.

[0109] 4.3 Strategy Solution and Output: The above model is solved by mixed integer linear programming (MILP) or simplex method, and the optimal oxygen supply regulation and power plant gas blending instructions for the next M time points are output and sent to the execution mechanism.

[0110] Specific process parameters and measures for oxygen regulation: In the strategy output and execution feedback module, the specific process parameters and control measures for oxygen regulation include: when multi-step rolling prediction results indicate a sharp drop in future converter gas production (e.g., a prediction that the converter will soon stop blowing gas), and the current oxygen pipeline pressure is too high, the system calculates the required oxygen reduction in advance and issues commands through the DCS system to reduce the guide vane opening of the oxygen generator (to reduce the actual oxygen production), or opens the oxygen pipeline compressor bypass valve in advance and reduces the flow rate of the liquid oxygen vaporizer, thereby avoiding oxygen release caused by pipeline overpressure. Conversely, when a large flow of gas is predicted to be generated (converter starting blowing gas), the oxygen generator load is increased in advance, and the adjustment advance is set to ensure that the oxygen blowing process is not affected.

[0111] Detailed mechanism for final closed-loop feedback: The specific implementation mechanism of the system's closed-loop feedback is as follows: After issuing the aforementioned coordinated control strategy, the system collects the actual feedback curves of the actuator (such as the actual valve opening) status and pipeline pressure in real time. The actual pipeline pressure data is compared and calculated in real time with the theoretical pressure curve obtained from multi-step rolling prediction. If the absolute value of the prediction deviation exceeds a preset tolerance threshold (e.g., set to 5%) for N consecutive sampling periods, the system determines that the current prediction model has deviated or that a sudden operating condition has occurred on-site (such as a temporary equipment failure). At this time, the closed-loop feedback mechanism is triggered, and the system immediately terminates the current preset control strategy, forcibly extracts the latest actual feedback data as a new query sequence, re-executes historical sample similarity matching and multi-step rolling prediction, and quickly generates new corrective control commands, thereby ensuring the dynamic robustness of the entire coordinated control system under strong disturbance conditions.

[0112] Application Scenario Demonstration Example: To further illustrate the control logic of the method of this invention in actual production, a typical production period (peak electricity price period) is selected for simulation. Taking the cluster of three 150-ton converters in this embodiment as an example, the three converters operate in parallel and share the same gas recovery main pipeline network. In actual production, overlapping operating conditions of two or even three converters cross-blowing often occur. Traditional mechanism models have great difficulty in accurately calculating the fluid coupling effects at multiple nonlinear hysteresis sources and pipeline intersections when facing overlapping gas production from multiple furnaces.

[0113] The sliding window matrix constructed in this invention not only includes the state of a single converter but also encompasses the real-time global collaborative formation of three converters. During the trial operation phase, the system automatically and implicitly extracts the 'comprehensive physical lag constant' of the converter cluster under cross-operation conditions through statistical analysis of historical data from the entire plant. τ 综合Approximately 2.5 minutes. Therefore, the system sets the future prediction length M to 15 minutes. When complex oxygen blowing characteristics of multiple furnaces overlap at the current time t0, the system can accurately predict the actual comprehensive gas volume peak that the total pipeline will experience after t0+2.5 minutes without any manual pipeline mapping or fluid dynamics modeling, through the implicit mapping of the algorithm. This completely overcomes the industry challenge of predicting gas volume in complex pipeline networks under multiple furnace cross-operations. Assuming the current time is... t 0, the system predicts the future M=15 minutes (i.e. t 0 to t 15 Rolling prediction and optimization are performed on the operating conditions of the machine.

[0114] 1. Prediction Phase: Using the similarity matching and multi-step rolling prediction mechanism described in steps three and four, the system identifies a high degree of similarity between the current converter operating condition (two of the three converters are in the middle of the blowing process) and "Sample Set #2024" in the historical sample library. The multi-step prediction sequence is ultimately generated through iterative rolling. Display: In the future to During this period, as both converters simultaneously enter the intense decarbonization phase, the gas production will reach a significant peak, with the total flow expected to reach 150,000 Nm³ / h (far exceeding the average).

[0115] 2. Optimization calculation stage: The predicted sequences were input into the oxygen dynamic regulation model.

[0116] Constraint Check: Model calculations revealed that without intervention, at the current consumption rate, the gas holder level... S t In t The upper limit was reached at 8 o'clock. S max This violates the constraints of formula (11).

[0117] Target optimization: The model is based on formula (14) (minimum discharge) and formula (15) (maximum benefit), at peak voltage ( Price peak Driven by incentives, they tend to consume surplus coal gas through power generation.

[0118] 3. Strategy Generation and Execution: The following comprehensive control strategy optimizes the solver output: Instruction A (Power Generation): In t Immediately issue a "load increase" pre-instruction to the self-owned power plant, requiring it to... t Five hours before the gas co-firing is increased to full load.

[0119] Instruction B (counter location): Int 1 to t During the fourth period, the opening of the gas holder inlet valve should be appropriately reduced to temporarily store some gas using the pipeline network's own capacity. t The flood peak five years later freed up cabinet space.

[0120] Instruction C (Oxygen): Predicted t 10 After the converter blowing process ends, oxygen demand will drop sharply. The system generates a "reduce oxygen generator load" suggestion in advance, planning to... t Start reducing the guide vane opening at 8 o'clock to avoid t 10 Oxygen is released at all times.

[0121] 4. Comparison of effects: After adopting the strategy of this invention, actual operation data shows that the gas venting rate was 0% during this period, and an additional 2000 kWh of electricity was generated during peak electricity price periods. Simultaneously, the standard deviation of the oxygen main pressure fluctuation decreased by 15%, effectively avoiding frequent speed adjustments of the oxygen generator unit caused by severe pressure fluctuations. In contrast, if traditional manual scheduling were used (which typically detects cabinet alarms 5-10 minutes later), it would inevitably lead to approximately 5000 Nm 3 The gas was released.

[0122] This embodiment also provides a system for implementing the above-mentioned method for synergistic optimization of converter gas and oxygen. For example... Figure 6 As shown, this system is deployed in the energy management center server cluster of a steel company and connects to the underlying distributed control system (DCS) and manufacturing execution system (MES) via industrial Ethernet. The system includes the following core functional modules: (1) Data Acquisition and Interface Module: This module is used to establish a real-time communication link with the production site. The module reads the instantaneous values ​​of converter gas pipeline pressure, oxygen main pipeline pressure, and flow meter readings at 1-second intervals via the OPC UA protocol; and synchronously obtains smelting furnace number, blowing status, and raw material structure data from the MES system through an intermediate database interface.

[0123] (2) Data preprocessing and storage module: This module is used to perform data cleaning and persistence. The module has a built-in algorithm processing unit, configured as follows: Missing value imputation: Using formula (1), interpolation is performed to repair data gaps caused by transmission interruption based on the K-nearest neighbor algorithm; Perform anomaly detection: Using formulas (2) and (3), identify and remove outliers caused by sensor drift based on the box plot method; Perform filtering and smoothing: Use formula (4) to perform median filtering on the high-frequency fluctuating pressure signal to extract the true trend.

[0124] The processed, standardized data is written into a time-series database (such as InfluxDB) for storage.

[0125] (3) Historical Trend Sample Set Management Module: This module is used to maintain the knowledge base required for similarity matching. The module is configured to slice historical data using a sliding window technique based on a set historical length L and future length M. For each time window, the module constructs a historical comparison sequence matrix according to formula (6). P i Construct the future prediction sequence vector based on formula (7). Q i And store the two together as a sample pair { P i , Q i This module uses an asynchronous thread to automatically update the sample set every day at midnight.

[0126] (4) Online prediction and optimization engine module: This module is the core computing unit of the system and is deployed on a high-performance computing server. The module is configured to perform the following operations: Real-time prediction: The latest real-time data sequence is loaded every minute, and its Manhattan distance with each sequence in the sample library is calculated using formula (8). Then, weighted fusion is performed using formulas (9) and (10) to generate a single-step prediction value. Subsequently, a system state prediction sequence for the next M steps is generated through an iterative rolling mechanism. V prod .

[0127] Collaborative optimization: Load the oxygen dynamic control model, which includes the cabinet position and pressure safety constraints defined by formulas (11) and (12). The module calls the mixed integer linear programming (MILP) solver to optimize the bi-objective function defined by formula (14) (minimize venting) and formula (15) (maximize benefit) under the premise of satisfying the constraints.

[0128] (5) Strategy Output and Execution Feedback Module: This module is used for closed-loop control. The module parses the optimal solution vector U output by the optimization engine into specific control commands (such as the oxygen generator guide vane opening setpoint and the power plant boiler gas valve opening), and sends them to the actuators through the DCS interface. At the same time, the module continuously monitors the pipeline pressure feedback after the command is executed. If the prediction deviation exceeds a preset threshold (such as 5%), an alarm is triggered and the optimization calculation is restarted.

[0129] The computer device in this embodiment includes a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, it performs the data cleaning, similarity matching, rolling prediction, and optimized scheduling steps described above. Experiments show that, compared to traditional manual scheduling or single PID control, this system reduces the converter gas emission rate by approximately 15% and the oxygen emission rate by approximately 10%, significantly improving energy utilization efficiency.

Claims

1. A converter gas prediction method based on trend perception, characterized in that, include: Data acquisition and preprocessing: Collect coal gas data, oxygen blowing data, and process parameters related to coal gas production, and preprocess the data to obtain a standardized historical dataset; Dynamic extraction of the comprehensive physical lag constant: Based on the historical dataset, calculate the maximum cross-correlation function between the total oxygen blowing flow sequence and the total gas recovery flow sequence of multiple furnaces, and determine the delay step size corresponding to the extreme point of the maximum cross-correlation function as the comprehensive physical lag constant of the current converter cluster. ; Adaptive Prediction Window Setting: Set the Length of Historical Trend Comparison And set the future prediction length. Such that the future prediction length The corresponding prediction duration is greater than or equal to the composite physical lag constant. ; Constructing a historical trend sample set: The standardized historical dataset is divided using a sliding window technique, and each window index is... , extract [ The data within the interval constitutes a historical comparison sequence, and the data is extracted from [...]. The data within the specified interval constitutes the future prediction sequence. All historical comparison sequences are associated with the future prediction sequence to form a historical trend sample set. Real-time single-shot prediction based on similarity matching: truncate the latest data from the end of the current historical dataset to a length of... The data is used to construct a real-time query sequence; the similarity distance between this real-time query sequence and all historical comparison sequences in the historical trend sample set is calculated, and the top... The future prediction sequences corresponding to the most similar historical comparison sequences are fused and calculated to output a single prediction result. Multi-step rolling prediction: The first predicted value from a single prediction is added as a "pseudo-actual value" to the end of the historical dataset, while the oldest data point at the beginning of the dataset is removed to update the dataset; the real-time single prediction is repeated using the updated dataset until a value of length is obtained. The multi-step prediction result sequence.

2. The converter gas prediction method according to claim 1, characterized in that, The formula for the maximum cross-correlation function between the total oxygen blowing flow rate sequence and the total gas recovery flow rate sequence of the multi-furnace system is: in, For the first The total oxygen blowing flow rate of all converters in the plant is monitored at all times. For delay Total gas recovery flow rate after time step The total length of the sample. This is the overall physical hysteresis constant.

3. The converter gas prediction method according to claim 1 or 2, characterized in that, When performing real-time single prediction based on similarity matching, the similarity distance is calculated using Manhattan distance or Euclidean distance. After selecting the top K most similar sequences, a weighted fusion algorithm based on similarity distance is used to calculate the predicted value, where the historical sequence with the smaller distance is assigned a larger weight.

4. The converter gas prediction method according to claim 1 or 2, characterized in that, The historical comparison sequence includes gas data, oxygen blowing data, and process parameters at various time points; the future prediction sequence only includes converter gas recovery data; and the historical trend sample set is updated offline periodically.

5. A method for synergistic optimization of converter gas and oxygen, characterized in that, Its features include: Obtain the sequence of multi-step prediction results obtained by the method described in any one of claims 1-4, and the comprehensive physical hysteresis constant of the current converter cluster. ; Cross-medium oxygen consumption feedforward mapping: based on the aforementioned integrated physical hysteresis constant The predicted gas production value at future moments within the forecast period is reverse-mapped to the predicted dynamic oxygen consumption of the converter at that future moment. Construct an oxygen dynamic control model: Using the oxygen supply regulation amount as the control variable, construct an oxygen dynamic control model that includes dynamic collaborative constraints on oxygen pipeline pressure; the dynamic collaborative constraints on oxygen pipeline pressure are: at any future time within the prediction period, based on the predicted value of the converter's dynamic oxygen consumption and the control variable, calculate the total oxygen pipeline pressure and restrict it to meet the safety boundary. Model Solving and Regulation: Under multiple constraints including gas user priority, gas storage location and gas pipeline pressure, the oxygen dynamic regulation model is solved with the objectives of minimizing gas emissions and maximizing gas utilization efficiency, and the oxygen supply regulation amount at each future time point is output.

6. The method for synergistic optimization of converter gas and oxygen according to claim 5, characterized in that, The specific logic of the dynamic collaborative constraint on oxygen pipeline pressure is as follows: When the dynamic oxygen consumption of the converter at a certain future moment is predicted to drop sharply based on the forecast value of gas production, the dynamic supply flow rate of the oxygen generator in the control variable is forcibly reduced in advance to prevent the pressure of the total oxygen pipeline network at that future moment from exceeding the upper limit of the safety boundary.

7. The method for synergistic optimization of converter gas and oxygen according to claim 5 or 6, characterized in that, The objective function of the oxygen dynamic regulation model includes: Objective 1: Minimize the amount of coal gas emitted, by adjusting the amount of coal gas consumed by the power plant and the changes in the coal gas holder level to bring the total amount of gas emitted by the system close to zero; Objective 2: Maximize the efficiency of coal gas utilization by increasing the amount of surplus coal gas used for power generation during peak hours of time-of-use electricity pricing and storing coal gas in gas holders during off-peak hours to maximize power generation revenue.

8. A converter gas prediction and oxygen co-optimization system, characterized in that, include: The data acquisition and preprocessing module is used to collect coal gas data, oxygen blowing data, and process parameters, and to preprocess them to obtain a standardized historical dataset. The hysteresis constant extraction module is used to calculate the maximum cross-correlation function between the total oxygen blowing flow rate and the total gas recovery flow rate based on the historical dataset, and dynamically determine the comprehensive physical hysteresis constant of the current converter cluster. ; Historical trend sample set management module: used to set the length of future predictions. and make it greater than or equal to And a historical trend sample set is constructed using sliding window technology; The real-time single prediction module is used to perform real-time single predictions based on similarity matching and output the single prediction results. The multi-step rolling prediction module is used to obtain a multi-step prediction result sequence of length M by iteratively updating the dataset with predicted values ​​as pseudo-actual values.

9. The converter gas prediction and oxygen co-optimization system according to claim 8, characterized in that, The system also includes an oxygen dynamic control module, which includes: Feedforward mapping unit, used for based on the comprehensive physical hysteresis constant The predicted value of gas production at future moments is mapped inversely to the predicted value of dynamic oxygen consumption of the converter. The collaborative constraint solving unit is used to introduce dynamic collaborative constraints on oxygen pipeline pressure based on the predicted oxygen consumption value into the oxygen dynamic regulation model, and solve for the output oxygen supply regulation amount under the condition of satisfying multi-objective optimization.

10. A computer-storable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it can implement the method as described in any one of claims 1-7.

Images