Running result early prediction method, device, equipment and storage medium

Abstract

The specification provides a running result early prediction method, device, equipment and storage medium, including: obtaining a full-process cascade model, the process flow of the process industrial system is divided into n process links, the full-process cascade model contains n sub-models cascaded in the order of the process flow, each sub-model corresponds to a process link, and is used for predicting at least part of the running result of the process link in the future window. For the first sub-model, based on at least part of the working parameters of the first process link collected in the observation window, the input tensor of the first sub-model is constructed and input into the first sub-model; and for the i-th sub-model, at least based on at least part of the working parameters of the i-th process link collected in the observation window and the running result predicted by the i-1-th sub-model, the input tensor of the i-th sub-model is constructed and input into the i-th sub-model.

Description

Technical Field

[0001] This specification relates to the field of industrial systems technology, and in particular to methods, apparatus, equipment and storage media for predicting operational results in advance. Background Technology

[0002] Process industry systems refer to industrial production and control systems that use continuously or semi-continuously flowing materials such as gases, liquids, powders, or bulk materials as their objects. Through physical, chemical, or biological transformation processes such as mixing, heating, reaction, phase change, separation, or heat transfer, the physicochemical properties of these materials are altered, ultimately producing new products. For example, specific examples of process industry systems include environmental protection and water treatment engineering systems, waste incineration systems, and so on.

[0003] Process industry systems are characterized by long and complex industrial processes, making traditional feedback control methods that rely on delayed adjustments based on the operational results of these systems often inadequate. Taking waste incineration systems as an example, when critical operational results (such as fluctuations in steam volume or excessive emissions of pollutants at the end of the process) have already become abnormal, issuing adjustment commands based on current observations often results in a significant time lag between control actions and operational responses, severely impacting production safety, stability, economic efficiency, and environmental compliance.

[0004] Therefore, it is necessary to make advance predictions of the future operating results of process industry systems. By predicting key operating results in advance for a period of time, the system and operators can intervene in advance and proactively eliminate system fluctuations.

[0005] In the field of advanced prediction solutions, the industry has proposed prediction methods based on deep learning models. However, when using deep learning models, the internal processing logic of these solutions relies entirely on the statistical correlation of the data. For example, taking a waste incineration system as an example, if the model's prediction target is the concentration of pollutants at the chimney outlet over a future period, existing solutions may only collect relevant operational data from the flue gas purification stage for prediction. However, pollutant emissions are not solely determined by end-of-pipe flue gas purification; they are first generated during combustion and then purified by units such as SNCR, SDA, and SCR before being emitted. The model struggles to explain the dynamic relationship between the pollutant concentrations at the boiler outlet and the chimney outlet, resulting in low prediction accuracy. Summary of the Invention

[0006] To overcome the problems existing in related technologies, this specification provides methods, apparatus, equipment and storage media for predicting operating results in advance.

[0007] According to a first aspect of the embodiments of this specification, a method for predicting operating results is provided, the method being applied to a process industry system, comprising: A full-process cascade model is obtained. The process flow of the process industry system is divided into n process steps. The full-process cascade model contains n sub-models cascaded in the order of the process flow. Each sub-model corresponds to a process step and is used to predict at least part of the operating results of the process step in the future window.

[0008] For the first sub-model, an input tensor for the first sub-model is constructed based on at least some of the working parameters of the first process step collected within the observation window, and then input into the first sub-model; and for the i-th sub-model, an input tensor for the i-th sub-model is constructed based on at least some of the working parameters of the i-th process step collected within the observation window and the running results predicted by the (i-1)-th sub-model, and then input into the i-th sub-model.

[0009] At least the prediction result of the nth sub-model is obtained as the target prediction object; where n is not less than 2 and i = 2, ..., n.

[0010] According to a second aspect of the embodiments of this specification, an apparatus for predicting operating results is provided, the apparatus being applied to a process industry system, comprising: The model acquisition module is used to acquire the full-process cascade model. The process flow of the process industry system is divided into n process steps. The full-process cascade model contains n sub-models cascaded in the process flow order. Each sub-model corresponds to a process step and is used to predict at least part of the operating results of the process step in the future window.

[0011] The model execution module is used to construct the input tensor of the first sub-model based on at least some of the working parameters of the first process step collected within the observation window, and input it into the first sub-model; and to construct the input tensor of the i-th sub-model based on at least some of the working parameters of the i-th process step collected within the observation window and the execution result predicted by the (i-1)-th sub-model, and input it into the i-th sub-model.

[0012] The result acquisition module is used to acquire at least the running result of the nth sub-model prediction as the target prediction object; where n is not less than 2 and i=2,...,n.

[0013] According to a third aspect of the embodiments of this specification, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described in the first aspect.

[0014] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in the first aspect.

[0015] The technical solutions provided in the embodiments of this specification may include the following beneficial effects: In the embodiments of this specification, the process flow of the process industry system is divided into multiple process stages. The full-process cascade model includes multiple sub-models cascaded in the process flow sequence. Each sub-model corresponds to one process stage and is used to predict at least part of the operating results of that process stage within a future window. In addition to receiving the operating parameters of its own process stage, each sub-model also receives the operating results output by the previous sub-model.

[0016] As can be seen, firstly, the internal processing logic of the full-process cascade model designed in this scheme "restores" the working process of the process flow, and can explain the complex evolution of the intermediate process state of the system. The input of each sub-model depends on the prediction result of the previous stage, so that the sub-model of this stage can predict the future operation result of the next stage based on the future operation result of the previous stage. This enables the model to deduce the change trajectory of the future target operation result according to the operation to be executed in a real control scenario, thereby improving the prediction accuracy.

[0017] Secondly, each sub-model in the entire cascaded model is modularly designed, and the structure of each sub-model can be flexibly designed according to the operating mechanism of this process link, which can give full play to the advantages of various algorithms and improve the prediction accuracy of the target operation results.

[0018] Finally, the inputs and outputs of each sub-model within the full-process cascade model correspond to the actual process steps, allowing staff to understand the changes in each intermediate process step. The intermediate results have clear physical meaning, making the model more interpretable.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this specification. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this specification and, together with the description, serve to explain the principles of this specification.

[0021] Figure 1 This is a flowchart illustrating an exemplary method for predicting operational results.

[0022] Figure 2 This specification is a schematic diagram illustrating the boundary and key variables of a waste incineration system according to an exemplary embodiment.

[0023] Figure 3 This specification is a schematic diagram illustrating a process flow and a full-process cascade model coordination mechanism according to an exemplary embodiment.

[0024] Figure 4 This specification illustrates a three-layer cascaded model based on physical process decomposition according to an exemplary embodiment.

[0025] Figure 5 This is a schematic diagram illustrating the construction of an input tensor according to an exemplary embodiment.

[0026] Figure 6 This is a structural diagram of a Crossformer-based prediction model illustrated in this specification according to an exemplary embodiment.

[0027] Figure 7 This specification illustrates a flowchart of phased training, end-to-end serial fine-tuning, and online rolling prediction based on an exemplary embodiment.

[0028] Figure 8 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment of this specification.

[0029] Figure 9 This is a block diagram illustrating an apparatus for predicting operating results according to an exemplary embodiment. Detailed Implementation

[0030] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0031] The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of this specification. The singular forms “a,” “the,” and “the” as used in this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0032] It should be understood that although the terms first, second, third, etc., may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this specification, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0033] Process industry systems refer to industrial production and control systems that use continuously or semi-continuously flowing materials such as gases, liquids, powders, or bulk materials as their objects. Through physical, chemical, or biological transformation processes such as mixing, heating, reaction, phase change, separation, or heat transfer, the physicochemical properties of these materials are altered, ultimately producing new products. For example, specific examples of process industry systems include environmental protection and water treatment engineering systems, waste incineration systems, and so on.

[0034] Process industry systems are characterized by long and complex industrial processes, making traditional feedback control methods that rely on delayed adjustments based on the operational results of these systems often inadequate. Taking waste incineration systems as an example, when critical operational results (such as fluctuations in steam volume or excessive emissions of pollutants at the end of the process) have already become abnormal, issuing adjustment commands based on current observations often results in a significant time lag between control actions and operational responses, severely impacting production safety, stability, economic efficiency, and environmental compliance.

[0035] Therefore, it is necessary to make advance predictions of the future operating results of process industry systems. By predicting key operating results in advance for a period of time, the system and operators can intervene in advance and proactively eliminate system fluctuations.

[0036] In the relevant schemes for advanced prediction, prediction methods based on deep learning models have been proposed. However, when these schemes utilize deep learning models, the internal processing logic of the models relies entirely on the statistical correlation of the data. For example, taking a waste incineration system as an example, if the model's prediction target is the concentration of pollutants at the chimney outlet over a future period, existing schemes may only collect relevant operational data from the flue gas purification stage for prediction. However, pollutant emissions are not solely determined by end-of-pipe flue gas purification; they are first generated during combustion and then purified by units such as SNCR, SDA, and SCR before being emitted. The model struggles to explain the dynamic relationship between the pollutant concentrations at the boiler outlet and the chimney outlet, resulting in low prediction accuracy.

[0037] To address the aforementioned technical issues, this specification provides a method for predicting operational results in advance.

[0038] The embodiments described in this specification will now be described in detail.

[0039] like Figure 1 As shown, Figure 1 This is a flowchart illustrating an exemplary method for predicting operational results, applied to a process industry system, comprising steps 101-103: Step 101: Obtain the full-process cascade model. The process flow of the process industry system is divided into n process steps. The full-process cascade model contains n sub-models cascaded in the order of the process flow. Each sub-model corresponds to a process step and is used to predict at least part of the operating results of the process step in the future window.

[0040] Step 102: For the first sub-model, construct the input tensor of the first sub-model based on at least some of the operating parameters of the first process step collected within the observation window, and input it into the first sub-model. Also, for the i-th sub-model, construct the input tensor of the i-th sub-model based on at least some of the operating parameters of the i-th process step collected within the observation window and the predicted running results of the (i-1)-th sub-model, and input it into the i-th sub-model.

[0041] Step 103: Obtain at least the running result of the nth sub-model as the target prediction object; where n is not less than 2 and i = 2, ..., n.

[0042] Based on their role in the process flow system, this scheme categorizes variables into operational variables, environmental variables, and target variables. Operational variables are those that operators or the control system can directly adjust, such as the operating parameters of each process step. Environmental variables describe the system state and operational response path. Target variables characterize the operational effect of each process step.

[0043] Taking waste incineration systems as an example, such as Figure 2 As shown, the process flow of the waste incineration system is as follows: feeding and conveying multi-source waste, grate combustion and feeding / air distribution, waste heat utilization boiler / steam, flue gas purification SNCR / SDA / SCR, and chimney emission CEMS monitoring.

[0044] Operating variables include the motion parameters of the pusher and grate, the flow rates of primary air, secondary air and recirculated air, the segmentation ratio of primary air, the flow rates of SNCR and SCR ammonia water, the flow rate of SDA lime slurry, the flow rate of cooling water and related auxiliary media, etc.

[0045] Environmental variables include bed thickness, flue gas temperature above the drying grate, furnace temperature, oxygen content at economizer outlet, SCR catalyst bed temperature, and SDA outlet temperature.

[0046] The target variables include main steam flow rate, boiler outlet NOx concentration, boiler outlet SO2 concentration, chimney outlet NOx concentration, and chimney outlet SO2 concentration.

[0047] In this scheme, the operating variables refer to the various working parameters in the system, while the environmental variables and target variables refer to the operating results after the completion of this process step.

[0048] In industrial process systems, manipulated variables do not directly affect the final target variable. Taking a waste incineration system as an example, manipulated variables first influence target variables such as steam output and pollutant concentration by changing environmental variables such as furnace temperature, bed thickness, flue gas oxygen content, and reactor thermodynamic state. For instance, adjusting a manipulated variable like primary air volume will first alter environmental variables such as the oxygen and temperature fields in the combustion zone, and only then will it cause changes in main steam flow and pollutant generation levels. If the target variable is directly predicted using manipulated variables, the model often struggles to characterize the complex evolution of intermediate process states, leading to insufficient prediction accuracy under conditions of strong nonlinearity and time delay. Conversely, while using a large number of environmental variables directly as input features may improve fitting ability, it can cause the model to become overly reliant on current observations of the process state, weakening the identification of the mechanism of action of manipulated variables and resulting in unclear control meaning and insufficient interpretability during optimization.

[0049] Therefore, this scheme adopts a joint prediction approach. Each sub-model in this scheme, besides receiving controllable operating parameters as input, also receives the operational results of the previous process stage predicted by the sub-model of the previous process stage. The operational results reflect the environmental state or output level of that process stage. Using the operated variables as external drivers, and incorporating key environmental variables into the set of predictive variables, the evolution trends of environmental variables and target variables are jointly extrapolated in the same model. This allows for a more complete characterization of the operation-state-result response relationship while maintaining clear control logic.

[0050] As can be seen, firstly, the processing logic of the full-process cascade model designed in this scheme "restores" the working process of the process flow, and can explain the complex evolution of the intermediate process state of the system. The input of each sub-model depends on the prediction result of the previous stage, so that the sub-model of this stage can predict the future operation result of the next stage based on the future operation result of the previous stage. This enables the model to deduce the change trajectory of the future target operation result according to the operation to be performed in a real control scenario, thereby improving the prediction accuracy.

[0051] Secondly, each sub-model in the entire cascaded model is modularly designed, and the structure of each sub-model can be flexibly designed according to the operating mechanism of this process link, which can give full play to the advantages of various algorithms and improve the prediction accuracy of the target operation results.

[0052] Finally, the inputs and outputs of each sub-model within the full-process cascade model correspond to the actual process steps, allowing staff to understand the changes in each intermediate process step. The intermediate results have clear physical meaning, making the model more interpretable.

[0053] In process industry systems, the technological stages are interconnected, with the output of the upstream stage flowing into the downstream stage as input. This scheme can divide the process industry system into multiple technological stages based on differences in chemical reaction mechanisms and the environments in which they occur.

[0054] For example, taking environmental protection and water treatment engineering systems as an example, the process flow can be divided into three stages: the pre-anaerobic stage, the aerobic aeration stage, and the secondary sedimentation stage.

[0055] For example, taking a waste incineration system as an example, the main steam flow rate and the concentration of pollutants in the flue gas are the most representative control targets. Among them, the main steam flow rate reflects the level of waste heat recovery and power generation capacity, and its stable output near the rated operating conditions is the foundation for ensuring the system's thermal efficiency and improving operational economy.

[0056] Regarding pollutants, NOx and SO2 can be selected as the main targets for prediction. Due to the combined effects of the high heterogeneity of multi-source waste components, significant fluctuations in low heating value, and time-varying combustion state, the formation process of these two types of pollutants exhibits obvious nonlinear, highly volatile, and time-lag characteristics, which brings greater challenges to achieving emission standards and operational control.

[0057] Meanwhile, pollutant concentrations are not solely determined by the combustion process; the initial generation level is determined by the combustion process, while the final emission result is determined by the flue gas treatment process. Based on this sequential process logic, and considering that the on-site DCS system can simultaneously acquire online monitoring data of pollutants at the boiler outlet and chimney outlet, the waste incineration system's process flow can be divided into three stages: combustion energy recovery, initial pollutant generation, and flue gas purification. Each process stage corresponds to a sub-model, used for predicting combustion conditions, initial pollutant generation, and flue gas purification, respectively.

[0058] By breaking down the entire waste incineration process into three sub-models based on physical processes—combustion energy recovery, initial pollutant generation, and flue gas purification—the transmission logic from feed and air distribution operations to thermal state, pollutant generation, and end-of-pipe emissions can be clearly expressed, providing a clearer process explanation compared to a single black-box model.

[0059] It should be noted that the process steps in this scheme are not based on a one-to-one correspondence with the inherent process nodes of the process flow. For example, multiple process nodes can be grouped into the same process step. Alternatively, the aforementioned combustion energy recovery step and the initial pollutant generation step can be merged into one step, reducing the process flow from three steps to two.

[0060] Operating parameters refer to operational variables in a process. For example, they can be control signals issued by operators to a process industry system, reflecting the control behavior towards the system. Examples include material feed rate control, equipment motion control, and equipment environmental status control.

[0061] There are multiple adjustable operating parameters for each process step. This solution can select key operating parameters from each process step as the model input for that process step.

[0062] The operational result of each process step refers to the product produced by this process step and the environmental state of the equipment after operation, when this process step is driven by specific working parameters.

[0063] Taking waste incineration systems as an example, considering that many environmental variables in waste incineration systems have strong spatial correlations, such as temperature measurement points at different heights and locations in the furnace often having similar trends, including all of them in the model would significantly increase the dimensionality and bring about multicollinearity problems. Therefore, it is necessary to screen environmental variables in combination with process mechanisms and control requirements.

[0064] For example, the original process variable system includes 25 operational variables, 30 environmental variables, and 5 core target variables. Based on process correlation and predictive control requirements, this solution further selects 6 key environmental predictive variables from the 30 environmental variables, which, together with the 5 core target variables, constitute 11 predictive variables as the output results.

[0065] Among them, the five core target indicators cover two types of objects: steam and flue gas, including main steam flow rate, NOx / SO2 concentration at boiler outlet and NOx / SO2 concentration at chimney outlet; the six key environmental indicators include bed thickness, flue gas temperature above the drying grate, furnace temperature, oxygen content at boiler outlet, SCR catalyst temperature and SDA outlet temperature.

[0066] The above describes the total output of the three sub-models. Next, we will explain the input parameters and output results for each sub-model: Taking a waste incineration system as an example, the working parameters input to the sub-model corresponding to the combustion energy recovery stage may include primary air flow rate, secondary air flow rate, primary air section 1 flow rate, primary air section 2 flow rate, primary air section 3 flow rate, primary air section 4 flow rate, primary air section 5 flow rate, primary air section 6 flow rate, pusher position, drying grate position, combustion grate position, and burnout grate position.

[0067] The working parameters input to the sub-model corresponding to the original pollutant generation stage may include primary air flow rate, secondary air flow rate, primary air section 1 flow rate, primary air section 2 flow rate, primary air section 3 flow rate, primary air section 4 flow rate, primary air section 5 flow rate, primary air section 6 flow rate, pusher position, drying grate position, combustion grate position, and burnout grate position.

[0068] The working parameters input to the sub-model corresponding to the flue gas purification process may include SDA lime slurry flow rate, SDA cooling water flow rate, SCR ammonia water flow rate, SCR steam flow rate, primary air flow rate, recirculation air flow rate, upper flue gas temperature of the drying grate, and furnace temperature.

[0069] It should be noted that the working parameters input to the sub-model corresponding to each process step can include not only the working parameters of this process step, but also the working parameters of its upstream process steps.

[0070] If there are no operating parameters for the defined process step, such as if the entire process step is a biochemical reaction process and does not involve system or personnel control, the operating parameters of the upstream process step can be used as the operating parameters of this process step.

[0071] The sub-model predicts the operating results for the combustion energy recovery stage, including main steam flow rate, bed thickness, upper flue gas temperature of the drying grate, furnace temperature, and oxygen content at the boiler outlet. For example, the main steam flow rate is a product of the combustion energy recovery stage, while the bed thickness, upper flue gas temperature of the drying grate, furnace temperature, and oxygen content at the boiler outlet represent the environmental conditions of the equipment after the combustion energy recovery stage has been running.

[0072] The sub-models predict the operating results for the original pollutant generation stages, including the NOx concentration and SO2 concentration at the boiler outlet.

[0073] The sub-models predict the following operating results for the flue gas purification process: SCR catalyst temperature, SDA outlet temperature, NOx concentration at the chimney outlet, and SO2 concentration at the chimney outlet.

[0074] Operating parameters and actual operating results can be read from DCS (Distributed Control System) and CEMS (Continuous Emission Monitoring System).

[0075] Taking a waste incineration system as an example, in terms of monitoring operating parameters, the DCS system can be equipped with several sensor points to monitor key operating parameters such as temperature, pressure, flow rate, liquid level, current and valve opening, which can reflect the operating status of the waste incineration system under different operating conditions in a relatively complete manner.

[0076] For monitoring operational results, two CEMS systems can be configured. One system can be deployed on the vertical section of the chimney to monitor the clean flue gas status in real time after the entire purification process. The monitoring indicators include physical parameters such as flue gas flow rate, temperature, humidity, and oxygen content, as well as the concentrations of pollutants such as particulate matter, hydrogen chloride, SO2, NOx, carbon monoxide, and hydrogen fluoride. The other system can be deployed at the boiler outlet. This measuring point is located before the flue gas purification system and can directly reflect the original flue gas status and initial pollutant concentrations before end-of-pipe treatment.

[0077] In one embodiment, the observation window can capture a time interval up to and including the current moment, while the future window can capture a time interval following the current moment. In short, the predicted outcome of this scheme can be the system's trajectory over a future period.

[0078] Compared to recursive single-step prediction, which easily propagates the error from the previous step to subsequent time steps, causing a decline in prediction accuracy over a long time domain, this solution can output the running trajectory for a future period of time simultaneously, avoiding the problem of gradual accumulation of errors in recursive single-step prediction. It is more suitable for early warning and feedforward control of long-process and large-lag conditions in process industry systems.

[0079] like Figure 3 As shown, this specification further provides the coordination mechanism between the process flow and the full-process cascade model: Suppose the process flow of a process industry system is divided into n process steps. The full-process cascade model contains n sub-models cascaded in the order of the process flow. Each sub-model corresponds to one process step and is used to predict at least part of the operating results of that process step within a future window. For example, process step 1 corresponds to sub-model 1, process step 2 corresponds to sub-model 2, process step i corresponds to sub-model i, and process step n corresponds to sub-model n.

[0080] Sub-model 1 can take at least some of the operating parameters of process step 1 as input.

[0081] In addition to taking at least some of the operating parameters of process step 2 and the operating results predicted by sub-model 1 as inputs, sub-model 2 can also take at least some of the operating parameters of process step 1 as inputs.

[0082] In addition to taking at least some of the working parameters of process link i and the predicted operating results of sub-model i as inputs, sub-model i can also take at least some of the working parameters of at least some of the process links 1 to i-1 and at least some of the operating results of at least some of the sub-models 1 to i-1 as inputs.

[0083] Similarly, in addition to taking at least some of the working parameters of process link n and the predicted operating results of sub-model n as inputs, sub-model n can also take at least some of the working parameters of at least some of the process links 1 to n-1 and at least some of the operating results of at least some of the sub-models 1 to n-1 as inputs.

[0084] Of course, this scheme can use only the operational results predicted by the nth sub-model as the target prediction object. This target prediction object is used for online early warning, feedforward control, and multi-objective optimization regulation, while the operational results predicted by sub-models 1 to n-1 are only used as intermediate results within the model. Alternatively, the operational results predicted by sub-models 1 to n-1 can also be used as the target prediction object. For example, taking a waste incineration system as an example, the main steam flow predicted by the sub-model corresponding to the combustion energy recovery stage is also useful, as it can be used to reflect the energy recovery level of the system. Meanwhile, the NOx and SO2 concentrations at the chimney outlet predicted by the sub-model corresponding to the flue gas purification stage can reflect whether the emission quality meets the standards.

[0085] The above-mentioned target prediction objects have mature applications in online early warning, feedforward control, and multi-objective optimization regulation. This solution does not limit the specific application of target prediction objects in this area.

[0086] like Figure 4 As shown, taking the waste incineration system as an example, the full-process cascade model (referred to in the figure as the "three-level series prediction model based on physical process decomposition") includes three sub-models 1, 2 and 3.

[0087] Sub-model 1 is used to describe the impact of feeding and combustion operations on the energy recovery level. Its prediction target is the main steam flow rate, and it also predicts key thermodynamic state variables such as the flue gas temperature above the drying grate, the temperature above the furnace, and the oxygen content at the economizer outlet.

[0088] Sub-model 2 focuses on the direct impact of the combustion process on the initial formation of pollutants, predicting NOx and SO2 concentrations at the boiler outlet. Since pollutant formation is not only related to the manipulated variables but also highly dependent on the thermal state established in the previous stage, sub-model 2, in addition to including manipulated variables such as feed, grate motion, and air volume distribution, also incorporates SNCR ammonia flow rate and uses variables from the prediction results of sub-model 1 as input features to reflect the driving effect of the combustion thermal process on pollutant formation.

[0089] Sub-model 3 is used to characterize the flue gas treatment process in the denitrification and deacidification units and its impact on final emissions. Its prediction targets include SCR catalyst bed temperature, SDA outlet temperature, and NOx and SO2 concentrations at the chimney outlet. Sub-model 3 uses the boiler outlet pollutant concentration as the initial boundary condition before treatment, and introduces operational variables related to reagent dosage and auxiliary media in the flue gas treatment process to describe the dynamic response of the end-of-pipe purification process.

[0090] All three sub-models employ a multiple-input multiple-output (MIMO) structure, where multiple predictor variables share network parameters within the same stage and learn their joint dynamic evolution process within a unified feature space. This not only improves the efficiency of information utilization among different variables but also helps avoid physical inconsistencies that can occur when modeling them separately.

[0091] To ensure the effectiveness and accuracy of subsequent model training and validation, data preprocessing operations can be performed on the collected working parameters: Taking a waste incineration system as an example, second-level operational data can be obtained from the waste incineration plant's DCS and CEMS systems, eliminating abnormal operating periods such as shutdown, maintenance, start-up / shutdown transitions, and continuous missing values ​​of multiple variables. Scattered missing values ​​are filled using forward filling, linear interpolation, or conformal interpolation; outliers are identified and processed using physical boundary screening, local statistical judgment, and operational logic constraints.

[0092] This solution does not impose any restrictions on the data preprocessing method.

[0093] To improve the prediction accuracy of the model, this approach proposes unique processing methods for some working parameters before constructing the input tensor: In one embodiment, if the sampling frequency of any operating parameter is higher than a frequency threshold, the operating parameter is downsampled to obtain a processing result, which matches the operating cycle. The processing result is then used to construct the input tensor after reconstructing the operating parameter.

[0094] The frequency threshold can be in the second range, while the downsampling result can be in the minute range. The specific degree of downsampling can be matched with the operation cycle.

[0095] For example, taking a waste incineration system as an example, although a DCS can provide second-level monitoring data, it contains a large amount of high-frequency fluctuations caused by sensor noise, equipment inertia, and short-term disturbances. These signals cannot truly reflect the actual adjustment intentions of the operators. In contrast, the on-site adjustment cycle for key operational variables such as feeding, grate, air distribution, and reagent dosing typically occurs on a timescale of 15-30 minutes; the responses of steam flow, furnace temperature, and pollutant concentration to these adjustments are also more consistent with minute-level inertia and hysteresis characteristics.

[0096] Directly using a second-level model to guide upper-level control would not only introduce high-frequency noise, interfering with the discovery of the true impact of the manipulated variables, but could also lead to excessively frequent control commands, which is detrimental to the stability of engineering applications. Therefore, this embodiment downsamples the second-level monitoring data into a minute-level sequence, using the minute average as the model input, thereby better simulating the adjustment logic of manual operation and ensuring the stability of subsequent intelligent control strategies.

[0097] In one embodiment, the positions of the pusher and each grate section are operational variables with obvious periodic characteristics, and their monitoring signals typically manifest as sawtooth fluctuations formed by reciprocating motion. These position signals do not directly correspond to the actual feeding intensity or turning amplitude, but rather describe the instantaneous state of the operational action. Directly using the raw position signals not only makes it difficult to characterize the actual action amplitude set by the operator within a single action cycle, but also introduces a large amount of invalid high-frequency information.

[0098] To address this, this solution introduces "stroke characteristics" to characterize the actual range of motion set within a single motion cycle.

[0099] The action cycle of the position signal can be determined, and the stroke characteristics of the pusher or grate can be extracted based on the position signal within the action cycle. After reconstructing the position signal within the action cycle into stroke characteristics, it is used to construct the input tensor.

[0100] For example, by differentially analyzing the position signal, local extreme points can be identified, thereby dividing the independent pushing or grate movement cycles. For the first... Each cycle defines its journey. It is the difference between the maximum position and the starting position of the period within the period, as shown in formula (1).

[0101] in, Indicates the time of the pusher or grate. Location; Indicates the first The time interval of a complete action cycle; This is the starting time of the cycle.

[0102] In one embodiment, in processing air distribution-related variables, this scheme does not directly use the absolute values ​​of various air volumes, but instead constructs dimensionless proportional characteristics to represent the air distribution structure.

[0103] For example, the ratio between each type of air volume and the total air volume is determined, and the various types of air volumes are reconstructed to the corresponding ratios to construct the input tensor.

[0104] Specifically, air volume includes primary air, secondary air, and recirculated air. The primary air ratio is defined as the proportion of primary air flow to the total of primary and secondary air; the recirculated air ratio is defined as the proportion of recirculated air flow to the total of recirculated and secondary air, used to reflect the relative proportions between different air systems.

[0105] In addition, to describe the air distribution pattern along the length of the grate, the distribution ratio of primary air in different grate sections was further constructed, which is defined as the proportion of primary air volume in each grate section to the total primary air volume.

[0106] These proportional characteristics can eliminate the influence of absolute level fluctuations in air volume and more directly depict the operator's adjustment strategy for the air distribution structure under different operating conditions.

[0107] In one embodiment, existing black-box, all-variable modeling methods struggle to represent controllable operational paths. Environmental variables can reflect system state, but most are not directly adjustable controllable quantities by operators. If the model relies excessively on environmental variables that cannot be obtained in advance, it will weaken the identification of the action paths of operational variables, hindering subsequent optimization control.

[0108] In response, this scheme further improves the structure of the input tensor: When constructing the input tensor of the i-th sub-model based on at least some of the working parameters of the i-th process step collected within the observation window and the running results predicted by the (i-1)-th sub-model, an observation matrix and a future matrix can be constructed, and the observation matrix and the future matrix can be concatenated into the input tensor in time sequence.

[0109] The time dimensions of the observation matrix and the future matrix are the sizes of the observation window and the future window, respectively. The same feature dimension of the observation matrix and the future matrix corresponds to the same working parameters or running results.

[0110] The characteristic variables input to the observation matrix include the working parameters of the i-th process step and the running results of the (i-1)-th process step collected within the observation window.

[0111] The feature variables input to the future matrix include the operating parameters to be used in the i-th process step within the future window and the operating results of the i-1th process step predicted by the i-1th sub-model within the future window.

[0112] In this embodiment, the constructed input tensor simultaneously includes the historical and predicted future operating results of the previous process step, as well as the historical and future operating results of this step. During the prediction process, historical state information and known future control settings are used simultaneously, enabling the model to deduce the future target variable change trajectory based on the operation to be performed in a real control scenario.

[0113] In one embodiment, the feature variables input to the observation matrix may further include the operational results of the i-th process step collected within the observation window. The feature variables input to the future matrix may also include a mask of the operational results of the i-th process step within the future window.

[0114] The predicted operational outcome in any stage is observable in historical time but unknown in the future prediction time domain, and is the output object that the sub-model of any stage needs to deduce. The operating parameters, on the other hand, are known quantities throughout the entire time domain, and their future values ​​are given by the preset values ​​of the control system. Since the two types of variables have different knowability in the future time domain, this scheme adopts a joint sample construction method based on a masking strategy to solve the misalignment problem between the operating parameters and the operational outcome in the time dimension.

[0115] like Figure 5 As shown, the observation window length is set to... The length of the future window is For any sampling time First, extract the most recent Construct an observation matrix using all variable observations at each time step. Simultaneously extract the future Utilize variable information at each time step to construct the future matrix. .in, Indicates the feature dimension.

[0116] In the future matrix, the channels corresponding to the operating parameters are filled with the actual future setpoints to represent subsequent control actions; and the channels corresponding to the output results of the previous stage are filled with predicted values ​​to represent the subsequent results of the previous stage. Since the true values ​​of the channels corresponding to the predicted results of this stage are unknown, a zero-fill masking strategy is used for masking.

[0117] Finally, the observation matrix and the future matrix are concatenated along the time dimension to form a matrix with a total length of [value missing]. joint input tensor Therefore, the model output is defined as the future. The true evolution sequence of the results within each time step.

[0118] This structure allows historical state information and future operational information to participate in prediction within a unified time frame, thereby ensuring that the model extrapolates the trajectory of the target variable in the future time domain based solely on known historical observations and future control settings during the prediction process, which conforms to the real logic of industrial process control.

[0119] In one embodiment, for the n cascaded sub-models included in the full-process cascaded model, each sub-model can adopt model architectures such as Transformer, iTransformer, Informer, Autoformer, TCN (Temporal Convolutional Network), LSTM, and GRU to achieve prediction functionality. The model architectures of different sub-models can be the same or different. This specification does not impose any restrictions on the model structure of each sub-model.

[0120] In one embodiment, this solution proposes a model architecture based on an improved Crossformer, which can serve as the model architecture for any one or more sub-models in a full-process cascaded model.

[0121] For example, such as Figure 6 As shown, the structure of the model includes, in sequence, a data dimension segmentation module, a data embedding module, an encoding module, a decoding module, and a prediction output module.

[0122] In the data dimension segmentation module, the original long sequence of each feature variable in the input tensor can be divided into multiple independent sub-segments. For each segmented sub-segment, a linear transformation is performed using a shared projection matrix.

[0123] Although linear projection extracts the shape features of the fragments, since all fragments undergo the same matrix processing, the model itself cannot distinguish whether the feature vector comes from furnace temperature or primary air volume, nor can it distinguish whether it occurred 10 minutes ago or at the current moment.

[0124] To address this spatiotemporal localization problem, an explicit two-dimensional spatiotemporal location code is injected into each linearly transformed sub-segment in the data embedding module. This code consists of two parts: temporal location code and dimensional location code. The temporal location code marks the time index of the segment within the historical sequence. This allows the model to understand the chronological order of operational conditions and establish temporal causal logic. The dimensional location code marks the variable index to which the segment belongs. This enables the model to distinguish the physical properties of different sensors, even if their waveforms are similar, allowing the model to identify differences in their physical meaning.

[0125] The encoding module consists of several levels of encoders. Between adjacent levels, for the output sequence of the previous level, the input of the next level is generated by concatenating adjacent time steps and linear projection. This structure endows the model with multi-scale perception capabilities. The bottom sequence is the longest and has the highest resolution, which can be used to capture high-frequency and transient disturbances, preserving the microscopic details of the operating conditions. The top sequence is the shortest and has the largest sensing range, focusing on capturing macroscopic operating conditions and long-term trends.

[0126] The decoding module also consists of several layers of decoders. Each layer of decoder receives the output of the encoder at the corresponding layer and extracts the corresponding hidden layer prediction features through a two-stage attention mechanism. The hidden layer features are then mapped to the target prediction space via linear projection to generate a coarse-resolution prediction result. To fuse the prediction information from different layers, the decoding module introduces upsampling operators (such as linear interpolation) to restore the coarse-resolution prediction sequence to the original target resolution. Finally, the upsampling results from all layers are linearly superimposed in the prediction output module to obtain the final prediction output.

[0127] The two-stage attention mechanism logically decouples the computation of the time dimension from that of the variable dimension, sequentially executing cross-temporal self-attention and cross-dimensional router attention. The specific implementation is as follows: First, the input array of the data embedding module is sliced ​​along the dimension axis. For the first... Given a variable, extract its feature vector sequence for all time steps. The data is then processed using a multi-head self-attention mechanism. After attention extraction and layer normalization, the data is fed into a feedforward neural network for nonlinear mapping to analyze the extracted deep temporal features. Finally, multiple independent feature matrices containing temporal context information are concatenated along the dimension axis to reconstruct a feature array.

[0128] Secondly, the high-dimensional variable space is mapped to a low-dimensional latent variable space. The router vector is used as the query matrix, and the variable features are used as the key and value matrices. The router actively extracts key information from all variables and compresses and aggregates it into a global summary. Then, the original variable features are used as the query matrix, and the generated global summary is used as the key and value matrices. Each variable queries the global state according to its own needs and uses the global information to correct its own feature representation.

[0129] To address the coexistence of temporal evolution patterns and multiphysics coupling in process industry systems, a two-stage attention mechanism was designed for the model. This mechanism logically decouples the computation of the time dimension from the variable dimension, sequentially executing cross-temporal self-attention and cross-dimensional router attention.

[0130] In one embodiment, the input of the first sub-model mainly consists of the working parameters of this process step and does not receive prediction outputs from other sub-models. For downstream sub-models 2 to n, although historical real values ​​are used as input during their training phase, they must rely on the prediction results of upstream sub-models for actual simulation. If the independent sub-models are simply connected end-to-end to form a direct serial model, the prediction errors generated upstream will gradually accumulate during multi-stage propagation, leading to a significant decrease in the reliability of the prediction results at the end.

[0131] To address this, this manual provides a training method for a fully cascaded model: The j-th sub-model can be trained separately based on the error between the predicted operating result of the j-th sub-model and the actual operating result of the j-th process step; j=1,2,...,n.

[0132] If the training results of each sub-model meet the target, end-to-end training is performed on the entire cascaded model based on the error between the predicted running result of the nth sub-model and the actual running result of the nth process step. During end-to-end training, the ith sub-model receives the running result predicted by the (i-1)th sub-model.

[0133] For example, for the first sub-model, the input data during training can be the actual working parameters of this stage. Finally, the first sub-model is trained separately based on the error between the actual running result of this stage and the running result predicted by the first sub-model.

[0134] For the i-th sub-model, the input data during training can be the actual working parameters of the i-th stage and the actual running results of the (i-1)-th stage. Finally, the i-th sub-model is trained separately based on the error between the actual running results of the i-th stage and the running results of the i-th stage predicted by the i-th sub-model.

[0135] If the training performance of each sub-model meets the target, key metrics such as MAE, RMSE, R2, and MAPE can be used to measure the training effect. Similar to the inference phase, for the first sub-model, the operating parameters of this stage are input into it. For the i-th sub-model, the operating parameters of this stage and the predicted results of the (i-1)-th sub-model are input into it. Finally, based on the error between the predicted results of the n-th sub-model and the actual operating results of the n-th process stage, end-to-end training is performed on the entire cascaded model.

[0136] During training, a loss function can be constructed using multi-objective weighted mean square error, mean absolute error, or a combination of both, and weights can be set according to the importance, scale, and engineering constraints of different predictor variables.

[0137] In this embodiment, based on a fixed basic structure of each sub-model, further fine-tuning is performed in series to allow the downstream model to gradually adapt to the operating conditions using the upstream prediction results as input. This reduces the inconsistency between the training scenario and the application scenario, improving the coordination and practicality of the end-to-end simulation throughout the entire process. The resulting cascaded model retains the clarity of the mechanism expression in staged modeling while possessing the overall deductive capability for real-time optimization decisions, providing a simulation tool that better conforms to the actual operating logic for subsequent system regulation.

[0138] like Figure 7 As shown, taking a multi-source waste co-incineration production line as an example, the entire process of model training and deployment is introduced: First, second-level historical data were collected from DCS and CEMS. After removing shutdown, maintenance, and abnormal operation periods, the data was resampled to minute-level sequences. A full-process variable system was constructed based on the process flow, comprising 25 operational variables, 30 environmental variables, and 5 core target variables. Six key environmental variables were then selected as joint prediction targets. The prediction targets included main steam flow, boiler outlet NOx concentration, boiler outlet SO2 concentration, chimney outlet NOx concentration, and chimney outlet SO2 concentration.

[0139] During implementation, the positions of the pusher and grate are first converted from their original position values ​​into stroke characteristics, and the primary air, secondary air and recirculated air are converted into air volume ratios and segmented air distribution characteristics. Then, the input tensor is constructed according to the estimation window L and the prediction step size S, and the predicted running results are zeroed out in the channels of the future matrix.

[0140] Sub-model 1 takes the feed, grate, and air volume operating variables as inputs and outputs the main steam flow rate and key thermal conditions; Sub-model 2 takes the combustion operating variables, SNCR ammonia flow rate, and the output of sub-model 1 as inputs and outputs the NOx / SO2 concentration at the boiler outlet.

[0141] Sub-model 3 takes the boiler outlet pollutant concentration, flue gas treatment agent dosage and key state variables as inputs, and outputs the SCR catalyst temperature, SDA outlet temperature and chimney outlet NOx / SO2 concentration.

[0142] The three models are first trained independently, and then connected in series according to sub-model 1—sub-model 2—sub-model 3 and fine-tuned end-to-end.

[0143] In continuous deployment, the model reads the working parameters of the latest observation window and future window in each rolling cycle, and directly outputs the multivariate prediction curve within the future window (e.g., 1 to 15 minutes). For scenarios requiring high-precision real-time adjustment, a rolling prediction can be triggered every step; for scenarios requiring early warning of pollutant exceedance risks or prediction of material disturbances, the 15-step trend prediction results can be directly used as a basis for early intervention.

[0144] Corresponding to the embodiments of the foregoing methods, this specification also provides embodiments of the apparatus and the terminal to which it is applied.

[0145] Figure 8 This is a schematic diagram illustrating the structure of an electronic device according to an exemplary embodiment. Figure 8 As shown, at the hardware level, the electronic device 800 includes a processor 802, an internal bus 804, a network interface 806, a memory 808, and a non-volatile memory 810, and may also include other hardware required for business operations. One or more embodiments of this specification can be implemented in software, for example, the processor 802 reads the corresponding computer program from the non-volatile memory 810 into the memory 808 and then runs it. Of course, in addition to software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic module, but can also be hardware or logic devices.

[0146] Figure 9 This is a block diagram illustrating an apparatus for predicting operating results according to an exemplary embodiment. Figure 9 As shown, this device can be applied to, for example Figure 8 The electronic device 800 shown implements the technical solution of this specification. The device is applied in a process industry system and includes: The model acquisition module 902 is used to acquire the full-process cascade model. The process flow of the process industry system is divided into n process steps. The full-process cascade model contains n sub-models cascaded in the order of the process flow. Each sub-model corresponds to a process step and is used to predict at least part of the running results of the process step in the future window.

[0147] The model execution module 904 is used to construct the input tensor of the first sub-model based on at least some of the working parameters of the first process step collected within the observation window, and input it into the first sub-model; and to construct the input tensor of the i-th sub-model based on at least some of the working parameters of the i-th process step collected within the observation window and the execution result predicted by the (i-1)-th sub-model, and input it into the i-th sub-model.

[0148] The result acquisition module 906 is used to acquire at least the running result of the nth sub-model prediction as the target prediction object; where n is not less than 2 and i=2,...,n.

[0149] Optionally, the process industrial system is a waste incineration system, and the process flow is divided into a combustion energy recovery stage, a pollutant initial generation stage, and a flue gas purification stage.

[0150] Optionally, the operating results predicted by the sub-model corresponding to the combustion energy recovery stage include main steam flow rate, bed thickness, upper flue gas temperature of the drying grate, furnace temperature, and oxygen content at the boiler outlet; the operating results predicted by the sub-model corresponding to the original pollutant generation stage include NOx concentration at the boiler outlet and SO2 concentration at the boiler outlet; and the operating results predicted by the sub-model corresponding to the flue gas purification stage include SCR catalyst temperature, SDA outlet temperature, NOx concentration at the chimney outlet, and SO2 concentration at the chimney outlet.

[0151] Optionally, the device further includes a frequency reconstruction module, used to downsample any operating parameter if the acquisition frequency of any operating parameter is higher than a frequency threshold to obtain a processing result, the processing result being matched with the operation cycle; after reconstructing the any operating parameter into the processing result, it is used to construct an input tensor.

[0152] Optionally, the operating parameters include the position signal of the pusher or grate, and the device further includes a position reconstruction module for determining the action cycle of the position signal, extracting the stroke features of the pusher or grate based on the position signal within the action cycle, and reconstructing the position signal within the action cycle into the stroke features to construct the input tensor.

[0153] Optionally, the operating parameters include at least two types of air volume, and the device further includes an air volume reconstruction module for determining the ratio between each type of air volume and the total air volume, and reconstructing each type of air volume to the corresponding ratio to construct the input tensor.

[0154] Optionally, the model running module 904 is used to construct an observation matrix and a future matrix, and to concatenate the observation matrix and the future matrix in time sequence into an input tensor; wherein, the time dimension of the observation matrix and the future matrix are the size of the observation window and the future window, respectively, and the same feature dimension of the observation matrix and the future matrix corresponds to the same working parameters or running results; the feature variables input to the observation matrix include the working parameters of the i-th process step collected within the observation window and the running results of the (i-1)-th process step; the feature variables input to the future matrix include the working parameters to be used in the i-th process step within the future window and the running results of the (i-1)-th process step predicted by the (i-1)-th sub-model within the future window.

[0155] Optionally, the feature variables input to the observation matrix also include the running results of the i-th process step collected within the observation window; the feature variables input to the future matrix also include a mask of the running results of the i-th process step within the future window.

[0156] Optionally, the full-process cascaded model is trained as follows: based on the error between the predicted operating result of the j-th sub-model and the actual operating result of the j-th process step, the j-th sub-model is trained separately; j=1,2,...,n; when the training effect of each sub-model meets the standard, the full-process cascaded model is trained end-to-end based on the error between the predicted operating result of the n-th sub-model and the actual operating result of the n-th process step; wherein, during the end-to-end training, the i-th sub-model receives the predicted operating result of the (i-1)-th sub-model.

[0157] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0158] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the solution in this specification according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0159] This specification also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the aforementioned methods for predicting operating results provided in this application.

[0160] Specifically, computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks.

[0161] This specification also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of any of the aforementioned methods for predicting operating results in advance.

Claims

1. A method for predicting operational results in advance, characterized in that, The method is applied to process industry systems, including: Obtain a full-process cascade model. The process flow of the process industry system is divided into n process steps. The full-process cascade model contains n sub-models cascaded in the order of the process flow. Each sub-model corresponds to a process step and is used to predict at least part of the running results of the process step in the future window. For the first sub-model, an input tensor for the first sub-model is constructed based on at least some of the working parameters of the first process step collected within the observation window, and then input into the first sub-model; and for the i-th sub-model, an input tensor for the i-th sub-model is constructed based on at least some of the working parameters of the i-th process step collected within the observation window and the running results predicted by the (i-1)-th sub-model, and then input into the i-th sub-model. At least the prediction result of the nth sub-model is obtained as the target prediction object; where n is not less than 2 and i = 2, ..., n.

2. The method according to claim 1, characterized in that, The process industrial system is a waste incineration system, and the process flow is divided into a combustion energy recovery stage, a pollutant initial generation stage, and a flue gas purification stage.

3. The method according to claim 2, characterized in that, The operational results predicted by the sub-model corresponding to the combustion energy recovery process include main steam flow rate, bed thickness, upper flue gas temperature of the drying grate, furnace temperature, and oxygen content at the boiler outlet. The sub-model predicts the operating results of the original pollutant generation process, including the NOx concentration and SO2 concentration at the boiler outlet. The sub-model predicts the operating results of the flue gas purification process, including SCR catalyst temperature, SDA outlet temperature, NOx concentration at the chimney outlet, and SO2 concentration at the chimney outlet.

4. The method according to claim 1, characterized in that, The method further includes: If the acquisition frequency of any working parameter is higher than the frequency threshold, the working parameter is downsampled to obtain the processing result, which is matched with the operation cycle. After reconstructing any of the working parameters into the processing result, it is used to construct the input tensor.

5. The method according to claim 1, characterized in that, The operating parameters include the position signal of the pusher or grate, and the method further includes: The action cycle of the position signal is determined, the stroke characteristics of the pusher or grate are extracted based on the position signal within the action cycle, and the position signal within the action cycle is reconstructed into the stroke characteristics to construct the input tensor.

6. The method according to claim 1, characterized in that, The operating parameters include at least two types of air volume, and the method further includes: Determine the ratio between each type of air volume and the total air volume, and reconstruct each type of air volume to its corresponding ratio to construct the input tensor.

7. The method according to claim 1, characterized in that, The construction of the input tensor of the i-th sub-model, based at least some of the operating parameters of the i-th process step collected within the observation window and the running results predicted by the (i-1)-th sub-model, includes: Construct an observation matrix and a future matrix, and then concatenate the observation matrix and the future matrix in time sequence into an input tensor; Wherein, the time dimensions of the observation matrix and the future matrix are the sizes of the observation window and the future window, respectively, and the same feature dimension of the observation matrix and the future matrix corresponds to the same working parameters or running results; The feature variables input to the observation matrix include the working parameters of the i-th process step and the running results of the (i-1)-th process step collected within the observation window; The feature variables input to the future matrix include the operating parameters to be used in the i-th process step within the future window and the operating results of the i-1th process step predicted by the i-1th sub-model within the future window.

8. The method according to claim 7, characterized in that, The feature variables input to the observation matrix also include the running results of the i-th process step collected within the observation window; The feature variables of the future matrix input also include the mask of the running result of the i-th process step within the future window.

9. The method according to claim 1, characterized in that, The full-process cascaded model is trained in the following way: Based on the error between the predicted operating result of the j-th sub-model and the actual operating result of the j-th process step, train the j-th sub-model separately; j=1,2,...,n; If the training effect of each sub-model meets the standard, end-to-end training is performed on the whole process cascade model based on the error between the running result predicted by the nth sub-model and the actual running result of the nth process step; wherein, during the end-to-end training, the ith sub-model receives the running result predicted by the (i-1)th sub-model.

10. A device for predicting operating results in advance, characterized in that, The device is used in process industry systems and includes: The model acquisition module is used to acquire the full-process cascade model. The process flow of the process industry system is divided into n process steps. The full-process cascade model contains n sub-models cascaded in the process flow order. Each sub-model corresponds to a process step and is used to predict at least part of the running results of the process step in the future window. The model execution module is used to construct the input tensor of the first sub-model based on at least some of the working parameters of the first process step collected within the observation window, and input it into the first sub-model; and to construct the input tensor of the i-th sub-model based on at least some of the working parameters of the i-th process step collected within the observation window and the execution result predicted by the (i-1)-th sub-model, and input it into the i-th sub-model. The result acquisition module is used to acquire at least the running result of the nth sub-model prediction as the target prediction object; where n is not less than 2 and i=2,...,n.

11. 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 program, it implements the steps of the method as described in any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1-9.

Images