A dry quenching working condition index prediction method and related device

Abstract

The application belongs to a working condition prediction method, aiming at the problem that the existing dry quenching working condition index prediction method is difficult to clearly predict the input and output boundary of the model based on material balance relationship and energy balance relationship, and is difficult to represent multivariate coupling and lag correlation, resulting in insufficient prediction accuracy of key process indicators at the target time, a method and related device are provided, the production boundary is determined based on the material balance relationship and energy balance relationship of the dry quenching system, and the input variables and output indicators associated with the to-be-predicted working condition index are determined within the production boundary, the input variables in the continuous historical time window are used to construct multi-dimensional time series input data; after standardization processing, input the working condition prediction model based on Transformer, extract the coupling relationship between variables and the cross-time step lag correlation relationship, and output the dry quenching working condition index prediction value at the to-be-predicted time. Therefore, the accuracy and stability of the key working condition index prediction can be improved.

Description

Technical Field

[0001] This application pertains to a method for predicting operating conditions, specifically a method and apparatus for predicting dry quenching coke operating condition indicators. Background Technology

[0002] During the operation of a dry quenching system, operating parameters such as circulating air volume, boiler feedwater volume, desuperheating water flow rate, introduced air volume, nitrogen consumption, pre-storage chamber pressure, and coke discharge rate collectively affect key process indicators such as superheated steam evaporation rate, coke discharge temperature, boiler inlet temperature, and circulating gas composition. Because material transfer and energy conversion processes occur within the dry quenching system, different variables do not change in isolation but exhibit strong coupling relationships and a certain response lag. Therefore, when predicting key process indicators, it is necessary to reasonably determine the model input and output variables and utilize historical operating data to characterize the temporal correlation between variables.

[0003] Most existing methods for predicting dry quenching coke operating conditions directly utilize field-collected data to build predictive models. They typically fail to first determine the production boundary by considering the material and energy balance relationships of the dry quenching system, and do not fully integrate the physical properties of variables, process correlations, and input-output relationships with key process indicators. Furthermore, they fail to differentiate between input variables and predicted output indicators in the operating data, resulting in unclear input-output boundaries for the predictive model. Simultaneously, existing methods underutilize multivariate data within continuous historical time windows when constructing model inputs, making it difficult to fully utilize the correlation information between different variables within the same time step and the temporal lag information between different time steps, thus affecting the accuracy of the predicted results for key process indicators at the target time. Summary of the Invention

[0004] This application addresses the technical problem that existing dry quenching coke condition index prediction methods struggle to clearly define the input and output boundaries of the prediction model based on material and energy balance relationships, and fail to fully utilize variable correlations and time lag information within continuous historical time windows, resulting in insufficient accuracy in predicting key process indicators at the target time. The application provides a dry quenching coke condition index prediction method and related apparatus.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application proposes a method for predicting indicators of dry quenching coke production conditions, including: The system acquires operational data of the dry quenching system within a continuous historical time window prior to the time to be predicted, and constructs multidimensional time series input data based on the input variables in the operational data; wherein, the input variables are operational variables associated with the operating condition indicators to be predicted within the production boundary of the dry quenching system; the production boundary is determined based on the material balance relationship and energy balance relationship of the dry quenching system. The multidimensional time series input data is standardized. The standardized multidimensional time series input data is input into a pre-trained Transformer-based operating condition prediction model. The lagged correlations of each input variable within the same time step are extracted, and the predicted values ​​of the dry quenching operating condition indicators at the time to be predicted are output. The Transformer-based operating condition prediction model is used to characterize the mapping relationship between the input variables within a continuous historical time window and the dry quenching operating condition indicators at the time to be predicted.

[0006] Furthermore, the material balance relationship includes the material conservation relationship corresponding to the entry of red coke, the discharge of cooled coke, the flow of circulating gas, and the boiler steam-water conversion; The energy balance relationship includes the energy conservation relationship corresponding to the sensible heat input of red coke, heat exchange of circulating gas, waste heat recovery of boiler, and steam output.

[0007] Furthermore, the input variables include at least one of the following: circulating air volume, boiler feedwater volume, desuperheating water flow rate, introduced air volume, nitrogen consumption, pre-storage chamber pressure, coke discharge volume, steam drum pressure, steam drum liquid level, boiler inlet circulating gas pressure, red coke charge volume, dry quenching furnace material level, coke powder temperature in the ash hopper under primary dust removal, dry quenching furnace inlet temperature, dry quenching furnace pre-storage chamber temperature, hydrogen content, carbon monoxide content, oxygen content, carbon dioxide content, circulating fan temperature, circulating fan inverter current, superheated steam evaporation volume, superheated steam temperature, coke discharge temperature, and boiler inlet temperature.

[0008] Furthermore, the predicted values ​​of the dry quenching coke condition indicators include at least one of the following: predicted value of superheated steam evaporation, predicted value of coke discharge temperature, predicted value of boiler inlet temperature, predicted value of hydrogen content, predicted value of carbon monoxide content, and predicted value of oxygen content. Among them, the variables in the input variables that have the same name as the predicted value of the dry quenching coke condition index are historical collection values ​​before the time to be predicted.

[0009] Furthermore, the step of constructing multidimensional time series input data based on input variables in the operational data includes: According to the sampling time sequence, extract the running data of T consecutive time steps before the time to be predicted; Within each time step, the input variables of the corresponding time step are concatenated to form a single time step input vector; Arrange the single-time-step input vectors of T consecutive time steps in chronological order to obtain multidimensional time series input data.

[0010] Furthermore, the multidimensional time series input data is standardized, including: Based on the preset mean and preset standard deviation, each variable in the multidimensional time series input data is standardized to eliminate the influence of differences in the dimensions of different variables on the working condition prediction model.

[0011] Furthermore, the operating condition prediction model includes an embedding module, a position encoding module, a Transformer encoder module, and a fully connected output module; The embedding module is used to map the standardized multidimensional time series input data into a feature vector; The position encoding module is used to inject time sequence information into the feature vector; The Transformer encoder module is used to extract input variable association features and temporal association features from the feature vector containing injected temporal sequence information; The fully connected output module is used to generate predicted values ​​of dry quenching coke condition indicators based on the correlation features of the input variables and the temporal correlation features.

[0012] Furthermore, the Transformer encoder module includes a multi-head attention layer, a feedforward network layer, a residual connection layer, and a layer normalization layer; The multi-head attention layer is used to calculate attention weights between different time steps and between each input variable in multiple feature subspaces of the feature vector injected with temporal sequence information. The feedforward network layer is used to perform a nonlinear transformation on the attention weights output by the multi-head attention layer; The residual connection layer and the layer normalization layer are used to fuse and normalize the attention weights and the features after nonlinear transformation.

[0013] Secondly, this application proposes a dry quenching coke condition index prediction system, including: The input data construction module is used to acquire the operating data of the dry quenching system within a continuous historical time window before the time to be predicted, and construct multidimensional time series input data based on the input variables in the operating data; wherein, the input variables are operating variables associated with the operating condition indicators to be predicted within the production boundary of the dry quenching system; the production boundary is determined based on the material balance relationship and energy balance relationship of the dry quenching system; A standardization processing module is used to standardize the multidimensional time series input data; The prediction module is used to input standardized multidimensional time series input data into a pre-trained Transformer-based operating condition prediction model, extract the lag correlation between each input variable within the same time step, and output the predicted value of the dry quenching operating condition index at the time to be predicted. The Transformer-based operating condition prediction model is used to characterize the mapping relationship between the input variables within a continuous historical time window and the dry quenching operating condition index at the time to be predicted.

[0014] Thirdly, this application proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the above-mentioned method for predicting dry quenching coke condition indicators.

[0015] Compared with the prior art, this application has the following beneficial effects: This application proposes a method for predicting dry quenching coke operating condition indicators. It acquires operational data of the dry quenching system within a continuous historical time window prior to the predicted time, and determines the production boundary of the dry quenching system based on material and energy balance relationships. Within this production boundary, input variables associated with the predicted operating condition indicators are identified. This ensures that the input variables of the prediction model originate from the material transfer, energy conversion, and state response processes within the same production boundary, avoiding the problems of unclear input-output boundaries and lack of process mechanism support when directly using field-collected data for modeling. This results in a clear physical and technological correlation between the input variables of the prediction model and the predicted operating condition indicators. Furthermore, this application constructs the input variables within the continuous historical time window as multi-dimensional time series input data. This ensures that the model input not only includes the operating state at a single moment but also the dynamic evolution information over a period prior to the predicted time. This provides a data foundation for subsequent models to utilize the time lag information between different time steps, improving the difficulty in accurately predicting the target operating condition indicators of the dry quenching system due to the lag in the thermal process response. Furthermore, this application standardizes the multidimensional time series input data before inputting it into the operating condition prediction model. This reduces the impact of differences in the dimensions and numerical ranges of different variables on the model calculation, and avoids variables with large numerical scales from exerting unreasonable dominant effects during model training or prediction, thereby improving the stability of the model training and online prediction processes. Further, this application uses a pre-trained Transformer-based operating condition prediction model to process the standardized multidimensional time series input data. Utilizing the Transformer model's attention modeling capability for sequence data, it models the variable correlations and temporal correlations within continuous historical time windows. This allows for a more comprehensive adaptation to the multivariate strong coupling and large lag characteristics of the dry quenching system, improving the accuracy and stability of the predicted values ​​of the dry quenching operating condition indicators at the time to be predicted. In summary, this application, through the coordinated efforts of production boundary determination, input variable construction, continuous historical time window construction, standardization processing, and Transformer prediction, enables the prediction process of dry quenching coke operating conditions to simultaneously possess process mechanism constraints and data-driven modeling capabilities. This solves the problems of existing methods that struggle to clearly define the input and output boundaries of the prediction model and fail to fully utilize variable correlations and time-series lag information, resulting in insufficient prediction accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for predicting dry quenching coke operating conditions according to this application. Figure 2 This is a collaborative optimization framework diagram of the dry quenching system in the embodiments of this application; Figure 3(a) is a schematic diagram of the material balance in the dry quenching system in the embodiment of this application; Figure 3(b) is a schematic diagram of the energy balance in the dry quenching system in the embodiment of this application; Figure 4 This is a schematic diagram of the variable association of the dry quenching system in the embodiments of this application; Figure 5 This is a schematic diagram of the dry quenching coke condition index prediction system of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] Dry quenching systems are commonly used in coking production for cooling red-hot coke and recovering waste heat. They are crucial equipment for improving energy efficiency and reducing pollution emissions in the coking process. The basic process involves loading high-temperature red-hot coke into a dry quenching furnace, where circulating inert gas exchanges heat with the coke, lowering its temperature. Simultaneously, the recirculated gas, having absorbed heat, is sent to a waste heat boiler to generate steam. Compared to traditional wet quenching, dry quenching reduces water consumption and emissions of pollutants such as dust and phenol-cyanide wastewater. It also recovers the sensible heat of the red-hot coke for power generation or steam supply, making it widely used in integrated steel mills and independent coking plants. During dry quenching operation, operating parameters such as circulating air volume, boiler feedwater volume, desuperheating water flow rate, introduced air volume, nitrogen consumption, pre-storage chamber pressure, and coke discharge rate directly or indirectly affect key process indicators such as superheated steam evaporation, coke discharge temperature, boiler inlet temperature, and circulating gas composition. Among them, the superheated steam evaporation rate reflects the waste heat recovery capacity, the coke discharge temperature is related to the coke cooling effect, the boiler inlet temperature affects the safety of the boiler heating surface and the heat exchange efficiency, and the composition of the circulating gas is related to the system inerting degree and safe operating status. Therefore, accurate prediction of the above key process indicators is of great significance for the optimized control, energy saving and consumption reduction and safe and stable operation of the dry quenching coke system.

[0021] In actual production, the dry quenching system is not a simple single-input, single-output device, but rather involves multiple processes such as the falling of red-hot coke, the flow of circulating gas, gas-solid heat exchange, boiler evaporation heat exchange, gas composition adjustment, and pressure balance. Because material transfer and energy conversion occur simultaneously and influence each other, there are often significant coupling relationships between operating parameters. For example, changes in circulating air volume not only affect the heat exchange intensity within the dry quenching furnace but also alter the boiler inlet temperature and steam production; changes in coke discharge rate alter the heat load entering the system per unit time, thus affecting the coke discharge temperature and circulating gas temperature; and the amount of air introduced and nitrogen consumed changes the circulating gas composition and the system's safety status. Furthermore, the dry quenching system exhibits significant thermal inertia and material retention characteristics. Changes in a particular operating parameter often do not immediately reflect the target process indicators but gradually manifest after a certain time delay. Therefore, relying solely on current operating data for prediction is insufficient to fully reflect the correlations and delayed responses between variables within the system.

[0022] To address the prediction of key process parameters in dry quenching coke production, existing technologies typically employ data-driven prediction models built from on-site collected operational data. For example, historical data such as circulating air volume, coke discharge volume, and boiler feedwater volume are used to predict parameters like steam output and coke discharge temperature through regression models, machine learning models, or neural network models. The advantage of this approach is its ability to utilize long-term accumulated data from the production site, reducing reliance on complex mechanistic formulas and, to some extent, learning the nonlinear relationship between operating parameters and process parameters. Some methods further improve model training through data cleaning, feature selection, and normalization. Normalization involves converting data of different dimensions and numerical ranges to a similar scale to avoid excessive weighting of certain large numerical variables in model training. Other methods attempt to incorporate time series modeling, using historical operational data as model input to depict trends in operating conditions. Overall, existing prediction methods primarily focus on data acquisition, feature processing, and model structure optimization, aiming to improve the prediction accuracy of key process parameters by enhancing model fitting capabilities.

[0023] However, dry quenching systems have clearly defined material and energy balance relationships. If these relationships are not considered before modeling to determine the production boundary, the predictive model may easily include data that does not belong to the same operating boundary or has no direct physical connection, leading to unclear input-output relationships. Meanwhile, field operation data includes variables adjustable by operators or control systems, process variables reflecting system state, and process indicators that need to be predicted or evaluated. If variables are not categorized into decision variables, state variables, and output variables based on their controllability, physical properties, and correlation with key process indicators, the model may use outcome variables, weakly correlated variables, or out-of-bounds variables as input during training, thus weakening the reliability of the prediction results. Decision variables are those that can be actively adjusted by operation or control strategies, state variables are those that characterize the current operating state of the system, and output variables are the key indicators that need to be predicted or monitored. Furthermore, existing methods do not adequately utilize multivariate data within continuous historical time windows when constructing model inputs, making it difficult to simultaneously express the coupling relationships between different variables within the same time step, as well as the lag relationships between different time steps. This results in the model not fully depicting the dynamic changes in the dry quenching system, especially when there are significant fluctuations in operating conditions, load adjustments, or changes in gas composition. The accuracy and stability of the prediction of key process indicators at the target time still have room for improvement.

[0024] Based on the above, this application proposes a method and related apparatus for predicting dry quenching operating conditions, used to predict key operating conditions at the desired time during the operation of a dry quenching system. The following detailed description, in conjunction with embodiments and accompanying drawings, further illustrates this application.

[0025] like Figure 1 The diagram shown is a flowchart illustrating one method for predicting dry quenching coke operating conditions according to this application. It may include: S101, acquire the operating data of the dry quenching system within a continuous historical time window before the time to be predicted, and construct multidimensional time series input data based on the input variables in the operating data.

[0026] It should be noted that the production boundary is used to define the material transfer range, energy conversion range, and state response range involved in predicting the current operating condition indicators. Input variables are operational variables associated with the predicted operating condition indicators within the production boundary of the dry quenching system. These can include operational quantities that can be adjusted by the control system, operating system, or operators during on-site operation, as well as monitoring quantities reflecting the current thermal state, material state, gas state, or equipment state of the dry quenching system. By determining the production boundary based on material and energy balance relationships, and then determining the input variables and predicted operating condition indicators within the production boundary, the input variables and output indicators of the prediction model can be derived from the same process boundary, avoiding the inclusion of external variables lacking process relevance to the predicted operating condition indicators into the model input.

[0027] After determining the input variables, the input variables within consecutive historical time windows are collectively constructed into multidimensional time series input data. Specifically, within each historical time step, the input variables corresponding to that time step are organized into single-time-step input vectors. These single-time-step input vectors are then arranged in chronological order of sampling time to form multidimensional time series input data oriented towards the time to be predicted. Through this processing method, the multivariate states within the same time step and the historical evolution processes between different time steps can both be incorporated into the subsequent prediction model.

[0028] S102, standardize the multidimensional time series input data.

[0029] In practical applications, standardization can be applied to each variable dimension based on preset mean and preset standard deviation, so that input variables with different dimensions and numerical ranges are transformed to a scale suitable for model calculation, thereby reducing the interference of dimensional differences on prediction results and improving the stability of the pre-trained working condition prediction model in the online prediction stage.

[0030] S103 inputs the standardized multi-dimensional time series input data into the pre-trained Transformer-based working condition prediction model, extracts the lag correlation of each input variable within the same time step, and outputs the predicted value of the dry quenching working condition index at the time to be predicted.

[0031] Among them, the Transformer-based operating condition prediction model is used to characterize the mapping relationship between input variables and dry quenching operating condition indicators at the time to be predicted within a continuous historical time window. During the prediction process, the Transformer-based operating condition prediction model can use an attention mechanism to weighted model the variable correlation and temporal correlation in the input sequence, enabling the model to use historical operating state information before the time to be predicted to output the predicted value of the dry quenching operating condition indicators at the time to be predicted.

[0032] The method described in this application supports the input and output boundaries of the prediction model by the material balance and energy balance relationships within the production boundary of the dry quenching system. The input variables are derived from the operational data within the production boundary that have material transfer, energy conversion, or state response relationships with the indicators of the working condition to be predicted. The model can model the variable correlation and time series correlation within a continuous historical time window, thereby improving the accuracy of the prediction of key process indicators at the time to be predicted.

[0033] Furthermore, since this embodiment first determines the production boundary of the dry quenching system based on material balance and energy balance relationships, and then determines the input variables associated with the predicted operating conditions within this production boundary, the input variables of the prediction model can be derived from operational data that have material transfer, energy conversion, and state response relationships with the predicted operating conditions, reducing the interference of irrelevant variables or variables outside the boundary on the prediction results. Simultaneously, the continuous historical time window can preserve the dynamic evolution process of the dry quenching system before the predicted time. The Transformer-based operating condition prediction model can utilize variable correlation information within the same time step and temporal lag information between different time steps within the same model, thereby improving the prediction stability and accuracy of operating condition indicators such as superheated steam evaporation, coke discharge temperature, boiler inlet temperature, and circulating gas composition.

[0034] The present application will be described in more detail below through some more detailed embodiments.

[0035] like Figure 2The diagram shows a collaborative optimization framework for a dry quenching system. In some embodiments of this application, the method for predicting dry quenching operating condition indicators can be embedded in the collaborative optimization framework for the dry quenching system. In this framework, the material balance and energy balance relationships of the dry quenching system are used to define the production boundary and to determine the input variables associated with the operating condition indicators to be predicted within the production boundary, as well as the operating condition indicators to be predicted. Based on this production boundary, the operating data of the corresponding input variables can be obtained from a continuous historical time window before the time to be predicted, and multi-dimensional time series input data can be constructed. Then, a Transformer-based operating condition prediction model is used to model the variable associations and time series associations within the continuous historical time window to obtain the predicted values ​​of the dry quenching operating condition indicators at the time to be predicted. It should be noted that... Figure 2 The optimization or control of operating parameters involved are only used to illustrate the usability of the predicted operating condition index values ​​in the dry quenching system. The dry quenching operating condition index prediction method of this application mainly consists of the input construction of the prediction model, standardization processing, Transformer feature extraction, and output of the predicted operating condition index values. Whether to further perform parameter optimization or control adjustment can be determined by the external system according to the on-site requirements and does not constitute a necessary step of this application.

[0036] Figure 3(a) shows a material balance diagram in the dry quenching coke system. When determining the production boundary of the dry quenching coke system, the dry quenching furnace, dust collector, boiler, and circulating fan can be included within the same material circulation boundary, and the conservation relationships of red-hot coke, cooled coke, circulating gas, boiler steam and water, and coke powder can be described within this boundary. Specifically, the material inputs on the dry quenching furnace side include the amount of red-hot coke charged, the amount of circulating gas entering, the amount of coke reaction gas, and the amount of introduced air; the material outputs include the amount of cooled coke discharged, the amount of coke burned off, the amount of coke powder, the amount of circulating gas leaving, and the amount of gas consumed in the coke reaction. The dust collector is used to collect the coke powder carried by the circulating gas and to generate the amount of coke powder carried out by the gas. The material inputs on the boiler side include the boiler feedwater and desuperheating water flow rates; the material outputs include the amount of gas released, the amount of superheated steam evaporation, and the amount of wastewater discharged. Nitrogen can also be introduced on the circulating fan side to maintain the safe inertization state of the circulating gas system. By understanding the material pathways described above, we can clarify the correspondence between variables such as coke discharge rate, boiler feedwater rate, desuperheating water flow rate, air intake rate, and nitrogen consumption and the material conservation of the dry quenching system. This provides a process basis for subsequently determining the input variables associated with the predicted operating conditions.

[0037] In one specific example of this application, the material balance relationship can be expressed as: in, This refers to the amount of red coke to be loaded. This refers to the amount of charred residue discharged. This refers to the amount of coke loss upon burning. For coke powder quantity, The amount of circulating gas entering the dry quenching furnace, To introduce air volume, To determine the amount of nitrogen to be introduced, This represents the amount of gas generated during the coke reaction. The amount of circulating gas leaving the dry quenching furnace, This refers to the amount of gas emitted. This refers to the amount of gas consumed in the coking reaction. The amount of coke powder carried out by the circulating gas. For water supply, To reduce the flow rate of heated water, This refers to the amount of steam evaporation. This represents the amount of wastewater discharged. The above relationship is used to characterize the material input and output correspondence between coke, circulating gas, and boiler steam / water within the dry quenching production boundary.

[0038] Figure 3(b) illustrates the energy balance in the dry quenching coke system. The energy balance includes the energy conservation relationships corresponding to the sensible heat input from the red-hot coke, the heat exchange of the circulating gas, the boiler waste heat recovery, and the steam output. The sensible heat carried by the red-hot coke serves as the main heat input to the system. After absorbing heat from the coke in the dry quenching furnace, the circulating gas flows to the boiler. The boiler utilizes the waste heat in the circulating gas to generate steam, and the steam output constitutes the waste heat recovery result. Through this energy transfer path, process correlations can be established between indicators such as superheated steam evaporation rate, coke discharge temperature, boiler inlet temperature, and circulating gas composition and the input variables within the production boundary.

[0039] In one specific example of this application, the energy balance relationship can be expressed as: in, To introduce sensible heat into the red-brown caramel, Heat is introduced into the circulating gas. The heat of coke combustion, To bring out the sensible heat from the coke. The circulating gas carries away sensible heat as it leaves the dry quenching furnace. To account for heat loss in dry quenching furnaces, The coke powder carries away sensible heat. The circulating gas carries the sensible heat into the boiler. For the sensible heat of the water supply, To reduce the sensible heat of the water, For steam heat, The circulating gas carries away sensible heat as it leaves the boiler. For boiler heat loss, Heat is carried out during blowdown. The above relationship is used to characterize the energy input and energy output correspondence between the sensible heat of red-hot coke, circulating gas heat exchange, boiler waste heat recovery, and steam output.

[0040] Figure 4 shows a schematic diagram of variable correlation in the dry quenching coke system. In the variable correlation diagram, the variables marked in blue and green can both be used as input variables for the operating condition prediction model, while the variables marked in orange can be used as predicted output indicators. These variables are interconnected along the circulation path formed by the dry quenching furnace, dust collector, boiler, and circulating fan. Changes in any operating variable can be transmitted to multiple output indicators through circulating gas flow, coke heat exchange, boiler steam-water conversion, and changes in gas composition. Therefore, this application uses the input variables associated with the operating condition indicators to be predicted within the production boundary as model inputs within a continuous historical time window, so that the Transformer-based operating condition prediction model can model the variable correlations and time-series correlations within the same historical time window.

[0041] Specifically, the input variables may include at least one of the following: circulating air volume, boiler feedwater volume, desuperheating water flow rate, introduced air volume, nitrogen consumption, pre-storage chamber pressure, coke discharge volume, steam drum pressure, steam drum liquid level, boiler inlet circulating gas pressure, red coke charge volume, dry quenching furnace material level, coke powder temperature in the ash hopper under primary dust removal, dry quenching furnace inlet temperature, dry quenching furnace pre-storage chamber temperature, hydrogen content, carbon monoxide content, oxygen content, carbon dioxide content, circulating fan temperature, circulating fan frequency converter current, superheated steam evaporation rate, superheated steam temperature, coke discharge temperature, and boiler inlet temperature.

[0042] The aforementioned input variables reflect the operational regulation behavior of the dry quenching system, the coke charging and cooling status, the furnace material level and temperature status, the circulating gas composition status, the dust collector ash discharge status, the circulating fan operation status, and the boiler steam-water system operation status, respectively. These variables describe the state changes of the dry quenching system during coke cooling, gas circulation, waste heat recovery, and equipment operation. Based on the variable distribution shown in Figure 4, the circulating air volume, introduced air volume, nitrogen consumption, pre-storage chamber pressure, and coke discharge volume mainly affect the dry quenching furnace and circulating gas system. The boiler feedwater flow rate and desuperheating water flow rate mainly affect the boiler steam-water conversion process. The steam drum pressure, steam drum liquid level, boiler inlet circulating gas pressure, dry quenching furnace material level, gas composition, circulating fan parameters, and temperature parameters can reflect the operational state response within the production boundary.

[0043] The predicted values ​​for dry quenching coke operating conditions can include at least one of the following: predicted superheated steam evaporation rate, predicted coke discharge temperature, predicted boiler inlet temperature, predicted hydrogen content, predicted carbon monoxide content, and predicted oxygen content. Among these, superheated steam evaporation rate reflects waste heat recovery output; coke discharge temperature reflects coke cooling effect; boiler inlet temperature reflects the thermal state of the circulating gas entering the boiler; and hydrogen, carbon monoxide, and oxygen content reflect the composition of the circulating gas and the operating conditions related to coke burnout and safe operation. It should be noted that when the input variables include superheated steam evaporation rate, coke discharge temperature, boiler inlet temperature, hydrogen content, carbon monoxide content, or oxygen content, these input variables are historically collected values ​​prior to the time of prediction; the predicted values ​​for dry quenching coke operating conditions are the predicted output values ​​at the time of prediction, and the two are distinct in time.

[0044] By defining the input variables and predicted output indicators as described above, the input variables characterize the operating status and changes of the dry quenching system during coke cooling, circulating gas heat exchange, boiler steam-water conversion, and equipment operation. The predicted output indicators characterize the key operating conditions that need to be monitored at the time of prediction. Therefore, the model input no longer relies solely on a single operating parameter or a single state monitoring value. Instead, it incorporates multiple types of operating variables associated with the predicted operating condition indicators within the production boundary into the historical time series. This allows the prediction process to reflect the continuous transmission relationship between operational adjustments, system state responses, and changes in target indicators, thereby improving the existing prediction methods' inability to adapt to strongly coupled and large-lag operating conditions.

[0045] The time to be predicted is designated as the target time. Continuous samples are taken before the target time in chronological order of sampling time. T The data is processed at each time step. Within each time step, the input variables for that time step are concatenated to form a single-time-step input vector. If 25 input variables are used, the single-time-step input vector contains 25 variable dimensions, continuously... T The single-time-step input vectors of each time step are arranged in chronological order to form multidimensional time series input data. , , Let be the time step. The model output can be represented as... , These are used to correspond to the six output indicators for the time to be predicted.

[0046] In some embodiments of this application, each variable in the multidimensional time series input data is standardized separately. For the first... T The time step, the first j The original input values ​​of each variable The following standardized forms can be adopted:

[0047] in, For the standardized variable values, For the first j The preset mean for each variable. For the first j Each variable has a predefined standard deviation. Each variable is processed using its own mean and standard deviation, so that variables with different dimensions such as circulating air volume, temperature, pressure, gas content, and evaporation are included in the same model calculation scale, avoiding the unreasonable dominant influence of variables with large numerical ranges in model training and prediction.

[0048] The Transformer-based condition prediction model can include an embedding module, a position encoding module, a Transformer encoder module, and a fully connected output module. The embedding module maps the standardized multi-dimensional time-series input data into feature vectors, transforming the original condition variables into feature representations suitable for attention calculation. The position encoding module injects temporal sequence information into the feature vectors, enabling the model to distinguish the sequential relationship between different sampling time steps and adapt to the time lag characteristic of the dry quenching system's output response.

[0049] The Transformer encoder module extracts input variable correlation features and temporal correlation features from feature vectors infused with temporal sequence information. This module includes a multi-head attention layer, a feedforward network layer, a residual connection layer, and a layer normalization layer. The multi-head attention layer computes attention weights in parallel across multiple feature subspaces, enabling the model to utilize the correlation information between input variables within the same time step from different perspectives, and to leverage the long-range dependence of input variables on future performance indicators across different time steps. Its computation process can be represented as follows:

[0050] in, , and The first i The query matrix, key matrix, and value matrix corresponding to each attention head are obtained from the input sequence features through a linear transformation. The dimension corresponding to the key matrix. For the first i The output of each attention head, The number of attention heads is considered. Through the above calculations, the model can directly establish the correlation strength between any two time steps and assign higher weights to strongly correlated historical operating states.

[0051] The feedforward network layer performs nonlinear transformation and deep fusion on the features output by the multi-head attention layer, enabling the model to express the complex nonlinear mapping relationship between input variables and the predicted working condition indicators. The residual connection layer and layer normalization layer are used to fuse and normalize the attention features and the nonlinearly transformed features to alleviate the gradient vanishing problem during deep network training and accelerate model convergence. One computational form can be expressed as:

[0052] in, For feedforward network modules, For layer-level normalization, the operating condition prediction model can extract abstract features layer by layer from the correlation between local variables and global temporal dependencies through multi-layer stacked encoder modules. The fully connected output module generates predicted values ​​of dry quenching operating condition indicators based on the features output by the encoder modules. Since the multi-head attention layer can calculate attention weights in multiple feature subspaces, the correlation strength between variables such as circulating air volume, boiler feedwater volume, coke discharge volume, furnace temperature, gas composition, and steam parameters within the same historical time window can be expressed differentially. At the same time, the attention calculation between different time steps allows the model to directly utilize historical state information that is far from the time to be predicted but still has an impact on the target indicators. Compared with the prediction method that only recursively extrapolates historical information in a fixed order, this embodiment can more fully characterize the characteristics of variable response lag, cross-path influence, and synchronous changes of multiple indicators in the dry quenching system, thereby improving the reliability of the predicted values ​​of operating condition indicators at the time to be predicted.

[0053] In one embodiment of this application, the training process uses historical multidimensional time series input data and future key outputs as training samples, and the model is trained end-to-end by minimizing the loss function between the predicted and actual values. The mapping relationship learned by the model can be expressed as:

[0054] in, The predicted value output by the model. These are the model parameters. The Transformer network after training. As a predictive model for the operating conditions of the dry quenching coke system, the network receives standardized historical operating data during the online prediction phase and outputs predicted values ​​such as superheated steam evaporation, coke discharge temperature, boiler inlet temperature, hydrogen content, carbon monoxide content, and oxygen content at the time to be predicted.

[0055] In some embodiments of this application, the predicted values ​​of dry quenching coke operating conditions output by the Transformer-based operating condition prediction model can be used for on-site operation status prediction, reference for adjusting operating parameters, advance judgment of alarm thresholds, or data input to external control systems. These predicted values ​​characterize the trends in superheated steam evaporation, coke discharge temperature, boiler inlet temperature, and circulating gas composition that may occur at the predicted time within the current continuous historical time window. Whether further parameter adjustments are made by an external optimization module, control module, or manual operating system can be determined based on on-site application requirements and does not constitute a mandatory step in the operating condition prediction method of this application.

[0056] Optionally, when constructing the input data and predicted output values ​​of the operating condition prediction model, the validity of the data range entering the model can be verified by combining the material balance relationship, energy balance relationship, process limits, and on-site equipment hardware limits of the dry quenching system. As an example, the validity verification range of some variables in Table 1 can be used: Table 1. Scope of Validity Check for Some Variables

[0057] The ranges in Table 1 characterize the typical operating intervals of some variables in the dry quenching system within the corresponding production boundaries. They can serve as a reference for screening training data for the operating condition prediction model, judging the validity of online input data, or verifying the rationality of prediction results. It should be noted that the variable constraints shown in Table 1 can be adjusted based on on-site equipment hardware limits, process safety limits, control system settings, or historical operating data ranges. The constraints in Table 1 do not require that all variables be verified in every prediction process, nor do they limit the application to performing an operating parameter optimization process.

[0058] like Figure 5 The diagram shown is a schematic representation of the dry quenching coke condition index prediction system of this application, which may include: The input data construction module is used to acquire the operating data of the dry quenching system within a continuous historical time window before the time to be predicted, and construct multidimensional time series input data based on the input variables in the operating data; wherein, the input variables are operating variables associated with the operating condition indicators to be predicted within the production boundary of the dry quenching system; the production boundary is determined based on the material balance relationship and energy balance relationship of the dry quenching system; A standardization processing module is used to standardize the multidimensional time series input data; The prediction module is used to input standardized multidimensional time series input data into a pre-trained Transformer-based operating condition prediction model, extract the lag correlation between each input variable within the same time step, and output the predicted value of the dry quenching operating condition index at the time to be predicted. The Transformer-based operating condition prediction model is used to characterize the mapping relationship between the input variables within a continuous historical time window and the dry quenching operating condition index at the time to be predicted.

[0059] In one embodiment of this application, an electronic device may also be provided. This electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can complete the data acquisition, variable partitioning, time series construction, standardization processing, Transformer model prediction, and predicted value output processes in the aforementioned dry quenching coke condition index prediction method. This electronic device can be deployed in a dry quenching coke production control system, process data server, edge computing terminal, or industrial control computer to provide predicted data for on-site operation adjustments, operational status assessment, or subsequent optimization control.

[0060] In one application example of this application, the operational data comes from the actual dry quenching system of a coking plant in northern China. The data collection period is from March 1, 2025 to May 31, 2025, with a sampling interval of 10 minutes, resulting in a total of 13,248 data points. The data is divided into training and testing sets in an 8:2 ratio according to chronological order. The experimental hardware includes an Intel i7-14700F CPU, 32GB of RAM, and an NVIDIA RTX 4060 8G GPU; the software environment includes a 64-bit Windows 10 operating system, PyCharm 2024 development environment, PyTorch 1.10 deep learning framework, and PyThon 3.8 runtime environment.

[0061] During the training of the prediction model, the Adam optimizer can be used with an initial learning rate of 0.002. A cosine annealing learning rate decay strategy is employed, training for 100 epochs until the learning rate decays to 0.0001. The L1 loss function can be used, and the dataset batch size is set to 16. When constructing training samples using a sliding window, with a historical time step T of 12, 10586 training data sets and 2637 test data sets can be generated.

[0062] In evaluating predictive models, the mean absolute error (MAE) and the coefficient of determination (R²) can be used to assess the model's fit. The formula for calculating MAE is:

[0063] The formula for calculating R² is:

[0064] in, For the true value, These are the model's predicted values. The average of the true values. This represents the number of samples.

[0065] In the example of historical time step selection, the Transformer was used as the load condition prediction model, and predictions were made with step sizes of 6, 12, 18, 24, and 30, respectively. The prediction results under different step sizes are shown in Table 2 below: Table 2. Transformer prediction accuracy at different historical time steps

[0066] Since the amount of superheated steam evaporation is directly related to sensible heat recovery, and the MAE is the lowest and the R² is the highest when the step size is 12, the step size of 12 can be regarded as a preferred historical time window length.

[0067] Table 3 shows the prediction results of different prediction models when the historical step size is 12: Table 3. Prediction results of each model when the historical step size is 12.

[0068] The results in Table 3 show that the Transformer model can achieve better prediction accuracy while maintaining low computational cost, making it suitable for predicting online operating conditions of dry quenching coke.

[0069] As demonstrated by the above application examples, under the same dry quenching operation data conditions, the historical step size has a significant impact on the prediction results. Using a continuous historical time window can provide the model with information on the state evolution before the time to be predicted. In this example, the Transformer model achieves superior MAE and R² evaluation results with lower computational cost, indicating its suitability for characterizing multivariate coupling and long-range lag correlations in dry quenching systems. This application example shows that the prediction method described in this application can generate stable operational condition prediction outputs from actual dry quenching operation data, providing a data foundation for on-site operational status prediction and operational adjustment references, but does not limit this application to performing an operational parameter optimization process.

[0070] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting dry quenching coke operating conditions, characterized in that, include: The system acquires operational data of the dry quenching system within a continuous historical time window prior to the time to be predicted, and constructs multidimensional time series input data based on the input variables in the operational data; wherein, the input variables are operational variables associated with the operating condition indicators to be predicted within the production boundary of the dry quenching system; the production boundary is determined based on the material balance relationship and energy balance relationship of the dry quenching system. The multidimensional time series input data is standardized. The standardized multidimensional time series input data is input into a pre-trained Transformer-based operating condition prediction model. The lagged correlations of each input variable within the same time step are extracted, and the predicted values ​​of the dry quenching operating condition indicators at the time to be predicted are output. The Transformer-based operating condition prediction model is used to characterize the mapping relationship between the input variables within a continuous historical time window and the dry quenching operating condition indicators at the time to be predicted.

2. The method for predicting dry quenching coke operating conditions according to claim 1, characterized in that: The material balance relationship includes the material conservation relationship corresponding to the entry of red coke, the discharge of cooled coke, the flow of circulating gas, and the boiler steam-water conversion. The energy balance relationship includes the energy conservation relationship corresponding to the sensible heat input of red coke, heat exchange of circulating gas, waste heat recovery of boiler, and steam output.

3. The method for predicting dry quenching coke operating conditions according to claim 2, characterized in that: The input variables include at least one of the following: circulating air volume, boiler feedwater volume, desuperheating water flow rate, introduced air volume, nitrogen consumption, pre-storage chamber pressure, coke discharge volume, steam drum pressure, steam drum liquid level, boiler inlet circulating gas pressure, red coke charge volume, dry quenching furnace material level, coke powder temperature in the ash hopper under primary dust removal, dry quenching furnace inlet temperature, dry quenching furnace pre-storage chamber temperature, hydrogen content, carbon monoxide content, oxygen content, carbon dioxide content, circulating fan temperature, circulating fan inverter current, superheated steam evaporation rate, superheated steam temperature, coke discharge temperature, and boiler inlet temperature.

4. The method for predicting dry quenching coke operating conditions according to claim 3, characterized in that: The predicted values ​​of the dry quenching coke condition indicators include at least one of the following: predicted value of superheated steam evaporation, predicted value of coke discharge temperature, predicted value of boiler inlet temperature, predicted value of hydrogen content, predicted value of carbon monoxide content, and predicted value of oxygen content. Among them, the variables in the input variables that have the same name as the predicted value of the dry quenching coke condition index are historical collection values ​​before the time to be predicted.

5. The method for predicting dry quenching coke operating conditions according to claim 1, characterized in that, The construction of multidimensional time series input data based on input variables in the operational data includes: According to the sampling time sequence, extract the running data of T consecutive time steps before the time to be predicted; Within each time step, the input variables of the corresponding time step are concatenated to form a single time step input vector; Arrange the single-time-step input vectors of T consecutive time steps in chronological order to obtain multidimensional time series input data.

6. The method for predicting dry quenching coke operating conditions according to claim 1, characterized in that, The multidimensional time series input data is standardized, including: Based on the preset mean and preset standard deviation, each variable in the multidimensional time series input data is standardized to eliminate the influence of differences in the dimensions of different variables on the working condition prediction model.

7. The method for predicting dry quenching coke operating conditions according to claim 1, characterized in that, The operating condition prediction model includes an embedding module, a position encoding module, a Transformer encoder module, and a fully connected output module; The embedding module is used to map the standardized multidimensional time series input data into a feature vector; The position encoding module is used to inject time sequence information into the feature vector; The Transformer encoder module is used to extract input variable association features and temporal association features from the feature vector containing injected temporal sequence information; The fully connected output module is used to generate predicted values ​​of dry quenching coke condition indicators based on the correlation features of the input variables and the temporal correlation features.

8. The method for predicting dry quenching coke operating conditions according to claim 7, characterized in that, The Transformer encoder module includes a multi-head attention layer, a feedforward network layer, a residual connection layer, and a layer normalization layer; The multi-head attention layer is used to calculate attention weights between different time steps and between each input variable in multiple feature subspaces of the feature vector injected with temporal sequence information. The feedforward network layer is used to perform a nonlinear transformation on the attention weights output by the multi-head attention layer; The residual connection layer and the layer normalization layer are used to fuse and normalize the attention weights and the features after nonlinear transformation.

9. A dry quenching coke condition index prediction system, characterized in that, include: The input data construction module is used to acquire the operating data of the dry quenching system within a continuous historical time window before the time to be predicted, and construct multidimensional time series input data based on the input variables in the operating data; wherein, the input variables are operating variables associated with the operating condition indicators to be predicted within the production boundary of the dry quenching system; the production boundary is determined based on the material balance relationship and energy balance relationship of the dry quenching system; A standardization processing module is used to standardize the multidimensional time series input data; The prediction module is used to input standardized multidimensional time series input data into a pre-trained Transformer-based operating condition prediction model, extract the lag correlation between each input variable within the same time step, and output the predicted value of the dry quenching operating condition index at the time to be predicted. The Transformer-based operating condition prediction model is used to characterize the mapping relationship between the input variables within a continuous historical time window and the dry quenching operating condition index at the time to be predicted.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the dry quenching coke condition index prediction method as described in any one of claims 1-8.

Images