Parameter optimization method for biomass gas coupled coal-fired boiler based on digital twinning

Abstract

The application relates to the technical field of industrial control, and in particular to a parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twinning. The application proposes the following scheme: a digital twinning model containing a mechanism submodel and a bias correction submodel is constructed, biomass gas sensing data and boiler operation data are fused to form a working condition state vector, and emission prediction is realized based on the bias distribution of simulation results and measured results. On this basis, a distribution bias factor is introduced to represent the structural characteristics and reliability of multi-dimensional bias, an emission constraint boundary varying with working conditions is generated, and a rolling optimization solution is combined with a combustion safety boundary to realize the collaborative adjustment of the blending ratio and air distribution parameters, thereby improving the stability and adaptability of emission control.

Description

Technical Field

[0001] This application relates to the field of industrial control technology, and in particular to a parameter optimization method for biomass gas coupled coal-fired boilers based on digital twins. Background Technology

[0002] In existing technologies, emission control for biomass gas coupled combustion processes typically relies on online monitoring results of the main components, calorific value, and boiler operating parameters of the biomass gas, combined with empirical models or fixed rules to adjust the co-combustion ratio and air distribution method. However, in actual operation, the above emission control methods still have shortcomings in terms of stability and adaptability.

[0003] On the one hand, existing emission control methods often deviate from actual emission behavior when biomass gas quality fluctuates or operating conditions change, making it difficult to reflect emission trends in a timely manner and resulting in emission control lag. On the other hand, when emissions approach limits, existing technologies often rely on passive adjustment or overall tightening control strategies, which can easily lead to excessive adjustment amplitude, discontinuous operation, or forced reduction in the biomass gas co-firing ratio, affecting the continuity and economy of coupled combustion operation. In addition, existing technologies mainly deal with emission fluctuations using single indicators or simple statistical methods, which are difficult to effectively cope with emission anomalies caused by multi-variable linkages, making emission control unreliable under complex operating conditions.

[0004] To address the above issues, this application proposes a parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins. Summary of the Invention

[0005] The technical problem this application aims to solve is to address the shortcomings of existing technologies by providing a parameter optimization method for biomass gas coupled with coal-fired boilers based on digital twins. This method constructs a digital twin model including a mechanistic sub-model and a deviation correction sub-model, integrates biomass gas sensing data and boiler operating data to form an operating condition state vector, and achieves emission prediction based on the deviation distribution between simulation and measured results. Furthermore, a distribution deviation factor is introduced to characterize the structural features and reliability of multidimensional deviations, generating emission constraint boundaries that vary with operating conditions. This is then combined with combustion safety boundaries for rolling optimization, enabling coordinated adjustment of the co-firing ratio and air distribution parameters, thereby improving the stability and adaptability of emission control.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins is applied to a production management system. The production management system is configured with a digital twin model, which includes a mechanistic sub-model and a deviation correction sub-model. The mechanistic sub-model is used to characterize the mass conservation, energy conservation, and pollutant generation mechanisms in the coupled combustion process of the biomass gas-coal-fired boiler. The method includes:

[0008] Acquire sensing data of biomass gas and boiler operation data, and generate state vectors;

[0009] The state vector is mapped to the simulation results of the digital twin model to obtain the deviation distribution between the state variables and emissions under the corresponding operating conditions. The deviation correction sub-model is then updated based on the deviation distribution to generate emission predictions within a preset time domain.

[0010] Based on the emission prediction, a constrained optimization problem including emission limits and combustion safety boundaries is constructed and solved to obtain optimized control quantities, wherein the optimized control quantities include at least one of the following: blending ratio, total air volume, and staged air distribution.

[0011] The sensing data includes the flow rate, temperature, pressure, and lower heating value of biomass gas, as well as the concentrations of the main components corresponding to carbon monoxide, hydrogen, methane, and carbon dioxide. The boiler operation data includes unit load, coal feed rate, primary air volume, secondary air volume, total air volume, furnace outlet oxygen content, furnace negative pressure, main steam flow rate, main steam pressure, main steam temperature, and reheat steam temperature.

[0012] The generated state vector includes:

[0013] The biomass gas sensing data and the boiler operation data are processed for time synchronization, outlier removal, and dimensional unification.

[0014] Principal component analysis was used to process the processed sensing data and boiler operation data, and characteristic quantities were selected to characterize combustion boundary conditions, oxygen distribution in the furnace, and heat input characteristics.

[0015] The feature quantities are combined in chronological order to generate a state vector characterizing the current coupled combustion condition. The state vector includes variables reflecting the energy and composition characteristics of biomass gas and variables reflecting the boiler load and air distribution status.

[0016] The method further includes:

[0017] The target working condition boundary conditions are determined based on the state vector, and the initial simulation results corresponding to the target working condition boundary conditions are determined by combining the digital twin model.

[0018] Extract historical operating condition data corresponding to the target operating condition boundary conditions within a preset time window. The historical operating condition data includes at least boiler operation data and actual emission data corresponding to the historical operating conditions.

[0019] Calculate the deviation between the initial simulation results and the historical operating condition data to obtain the reference deviation;

[0020] The initial simulation results are corrected by weighted fusion of the reference deviation to obtain the simulation results of the digital twin model. The weighting coefficients are determined by the degree of difference between the boundary conditions of the target operating condition and the boundary conditions of the historical operating condition, and the degree of difference is calculated based on the lower heating value of biomass gas, the concentration of main components and the unit load.

[0021] The state vector is mapped to the simulation results of the digital twin model to obtain the deviation distribution between the state variables and emissions under the corresponding operating conditions, including:

[0022] Determine the set of mapping variables corresponding to the state vector, wherein the mapping variables include derived state variables for characterizing the coupling relationship between fuel and wind distribution and check emissions for characterizing the emission response;

[0023] The corresponding measured values ​​are obtained for each mapping variable in the set of mapping variables, and the measured values ​​are matched with the simulation results output by the digital twin model at the same time.

[0024] The difference between the matched measured values ​​and simulation results is calculated to obtain the deviation value corresponding to each mapping variable.

[0025] The deviation values ​​are summarized according to preset statistical rules to generate a deviation distribution that characterizes the deviation amplitude and fluctuation range under the target working condition. The statistical rules include at least one of sliding time window statistics and quantile statistics.

[0026] The deviation correction sub-model is used to generate emission constraint boundaries within a preset time domain based on the deviation distribution, and to generate emission predictions through a prediction algorithm, specifically including:

[0027] Based on the deviation distribution, the fluctuation range of each verification emission amount within the preset time domain is determined, and the fluctuation range is converted into a safety margin.

[0028] The safety margin is combined with the preset emission limit to obtain the emission upper limit that varies with the target operating condition, wherein the emission upper limit is less than or equal to the preset emission limit.

[0029] The emission upper limit is used as the emission constraint boundary, and the emission prediction within the preset time domain is generated by combining the emission constraint boundary with the prediction algorithm.

[0030] Based on the emission predictions, a constrained optimization problem is constructed, including emission limits and combustion safety boundaries, and then optimized and solved, including:

[0031] The weights of the objective function are determined by the analytic hierarchy process to obtain the optimized objective function, which is used to reduce the combustion regulation cost while satisfying the emission constraint boundary.

[0032] The emission prediction values ​​at each time point within the preset time domain and the corresponding emission constraint boundaries at each time point constitute the emission constraint conditions.

[0033] Based on the boiler operating data, the combustion safety boundary is determined and the combustion safety boundary is used as a safety constraint condition. The combustion safety boundary includes at least one of the following: the lower limit of oxygen content at the furnace outlet, the allowable range of negative pressure in the furnace, the upper limit of main steam temperature, and the upper limit of reheat steam temperature.

[0034] The optimized control quantity is determined as a decision variable and the range and rate of change of the decision variable are limited, wherein the decision variable includes at least one of the following: co-firing ratio, total air volume, and staged air distribution.

[0035] In each control cycle, the optimized control quantity of the previous control cycle is used as the initial value. Combining the emission constraints and the safety constraints, the decision variables are solved in a rolling manner through the optimization objective function to obtain the optimized control quantity of the current control cycle.

[0036] The step of calculating the difference between the matched measured values ​​and simulation results to obtain the deviation values ​​corresponding to each mapping variable includes:

[0037] Within a preset sliding time window, the deviation values ​​corresponding to each mapping variable are arranged into a multidimensional deviation sequence according to the time order, and a deviation dataset is constructed to characterize the time-varying features of the multidimensional deviation sequence.

[0038] The bias dataset is subjected to sparse factor separation processing, which decomposes the bias dataset into multiple latent factor time series and factor loadings corresponding to each mapping variable.

[0039] The biased dataset is reconstructed based on the latent factor time series and factor loadings to obtain the reconstructed bias, and the unexplained residual between the biased dataset and the reconstructed bias is calculated, wherein the unexplained residual is used to characterize the interpretability of the sparse factor separation to the source of bias.

[0040] Based on the contribution structure of the reconstruction bias to the corresponding bias values ​​of each mapping variable, and combined with the unexplained residuals, the distribution bias factor is obtained through confidence-weighted quantile statistics.

[0041] The deviation correction sub-model includes factor-wise correction of the fluctuation range based on the distribution deviation factor, specifically including:

[0042] Based on the distribution deviation factor, the fluctuation range of each verification emission within the preset time domain is determined, and the fluctuation range of each sub-factor is weighted according to the contribution weight corresponding to the distribution deviation factor to obtain the comprehensive fluctuation range.

[0043] The overall volatility range is reliably corrected based on the unexplained residuals, and the sub-factor volatility range is scaled based on the reliably corrected overall volatility range.

[0044] The sub-factor fluctuation range represents the fluctuation range of the verification emission deviation corresponding to a single implicit factor within the preset time domain. The fluctuation range is determined by the component contribution of the implicit factor time series and the factor loading to the reconstruction deviation.

[0045] Compared with the prior art, the beneficial effects of this application are:

[0046] This application introduces a digital twin framework combining mechanistic models and deviation correction to uniformly map online sensing data with actual operational behavior, enabling emission predictions to be dynamically updated in response to changes in operating conditions. Building upon this, by structurally representing multidimensional deviations and forming a distributed deviation factor, the uncertainty of emission fluctuations is introduced into the constraint construction process in a calculable form, giving the emission constraint boundary the ability to adaptively adjust with operating conditions. Furthermore, by combining combustion safety boundaries with rolling optimization solutions, the co-firing ratio and air distribution parameters are coordinated, effectively improving the stability and continuity of emission control, reducing the impact of overly conservative adjustments on operation, and thus balancing emission compliance, combustion safety, and operational feasibility under complex coupled operating conditions. Attached Figure Description

[0047] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0048] Figure 1 An exemplary application scenario diagram provided for an embodiment of this application;

[0049] Figure 2 A schematic diagram of a processor module provided in an embodiment of this application;

[0050] Figure 3 A flowchart illustrating the parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins, provided in an embodiment of this application.

[0051] Reference numerals: 100, Biomass gasification device; 101, Biomass gas pipeline; 102, Coal-fired boiler; 103, Flue gas pipeline; 104, Emission monitoring device; 105, Production management system; 1051, Processor; 10511, Data acquisition module; 10512, Model storage module; 10513, Emission prediction module; 10514, Optimization control module. Detailed Implementation

[0052] The technical solutions in 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.

[0053] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0054] The method described in this application is applicable to coupled combustion operation scenarios formed by the introduction of biomass gas in coal-fired units. It is particularly applicable to gasification systems where the syngas originates from agricultural and forestry waste, straw, and lignin-based raw materials, and is then purified and fed into pulverized coal boilers as a gaseous alternative fuel to achieve partial heat substitution.

[0055] In such engineering practices, the production management system can typically and reliably acquire syngas flow rate, temperature, pressure, lower heating value, and concentrations of major components such as carbon monoxide, hydrogen, methane, and carbon dioxide. It can also collect operational data from the boiler side, including unit load, coal feed rate, primary air volume, secondary air volume, total air volume, furnace outlet oxygen content, furnace negative pressure, and main / reheat steam parameters. Simultaneously, online monitoring of pollutant emissions can provide measured values ​​for verification emissions of nitrogen oxides, carbon monoxide, sulfur dioxide, and particulate matter. This data chain makes emission prediction and control based on online sensing engineering feasible and forms the typical deployment basis for the embodiments of this application.

[0056] It should be noted that, unlike conventional coal-fired operation, emission control in biomass gas coupled combustion is not solely influenced by the magnitude of heat input or the excess air coefficient. Due to the dispersed sources and significant differences in the physicochemical properties of the gasification feedstock, syngas quality drift inevitably occurs during long-term operation due to fluctuations in feedstock, load, and process parameters. Although the production management system can characterize the main components and calorific value online, it is often difficult to stably and accurately quantify tar, ammonia, hydrogen chloride, hydrogen sulfide, or other latent components prone to cross-interference online.

[0057] After entering the furnace, the aforementioned hidden components will cause phenomena such as sudden increases in carbon monoxide, baseline drift of nitrogen oxides, and abnormal increases in secondary pollutants by changing the local reducing atmosphere, affecting the free radical chain reaction pathway, and inducing fuel nitrogen conversion and side reactions, even when the blending ratio, total air volume, and staged air distribution have not changed significantly.

[0058] For those skilled in the art, this type of phenomenon has the typical characteristics of being observable but difficult to attribute:

[0059] Anomalies in emissions can be detected at the monitoring end, but they are difficult to explain by existing principal component-calorific value equivalent models. This leads to problems such as uncertain adjustment direction, slow convergence speed, or failure after adjustment in the actual implementation of control strategies, forming a window of emission control.

[0060] In existing engineering strategies, a common approach is to treat syngas as an equivalent combustible gas with a certain calorific value and main component ratio, and to follow the empirical air distribution and emission control logic of coal-fired units. This involves increasing the excess air coefficient, adjusting the secondary air ratio, or reducing the co-firing ratio to maintain emission compliance. Exemplary technologies can play a certain role when syngas quality is relatively stable, but when the mechanistic pathways caused by latent components change, the emission response to the controlled quantities exhibits strong nonlinearity and path dependence. Simply relying on end-of-pipe emission feedback often results in a significant lag, and a conservative approach can easily lead to passive compression of the co-firing ratio, increased combustion regulation costs, and a difficulty in balancing economic efficiency and carbon reduction benefits.

[0061] It is important to emphasize that within the short time window of a sudden change in syngas quality, the deviation is not reflected in a single indicator, but rather forms an intrinsically linked deviation among furnace outlet oxygen, carbon monoxide, verification emissions, and derived state quantities related to the air distribution coupling relationship. If only the deviation of a single variable is statistically amplified, it is easy to mix the deviations from different sources into the same error magnitude, thus failing to support the effective identification and targeted suppression of the influence of hidden components.

[0062] Based on the aforementioned understanding, the core logic of the method in this application does not rely on the direct online quantification of latent components. Instead, it utilizes multi-source online data that can be stably obtained from the production management system to match the simulation output with the measured output at the same time within the framework of a digital twin model, forming a multi-dimensional deviation sequence covering the coupling relationship between fuel and wind distribution and the emission response.

[0063] Furthermore, considering that the influence of latent components on the system is distributed in the form of traces among multidimensional deviations and presents a repeatable structural pattern, this application introduces sparse factor separation processing based on a sliding time window to decompose the multidimensional deviation sequence into multiple latent factor time series and their loading relationships on each mapping variable, thereby realizing an online characterization of the source of deviations; at the same time, by comparing the reconstructed deviations with the unexplained residuals, the current deviation is quantified as to whether it is within the interpretable range, thus providing a reliable basis for the generation of subsequent emission constraint boundaries.

[0064] It should be noted that the traces here can be understood as the linkage deviation characteristics formed by the differences between the measured values ​​of multiple observable quantities and the digital twin simulation values ​​within the same time window. The linkage deviation characteristics are manifested as the consistency or reversal of the deviations of furnace outlet oxygen, carbon monoxide, verification emissions, and fuel and air distribution coupled derived state quantities in terms of amplitude, phase sequence, and direction of coordinated change.

[0065] In other words, at the implementation level, this application expands the statistical distribution that can only see the magnitude of deviation in the traditional sense into a decomposed deviation representation that can express the contribution and credibility of the deviation source. This enables those skilled in the art to quantify the degree of risk based on observable traces when faced with emission mutations caused by the unmeasurability of hidden components, and to achieve stable control in the constraint optimization solution with a more reasonable safety margin and feasible region tightening strategy.

[0066] It should be understood that the method described in this application does not require the controlled object to have a specific burner type or fixed furnace partitioning, nor is it based on a specific raw material or a single gasification route. The applicability of the embodiments of this application can be demonstrated as long as the application scenario simultaneously meets at least one of the following typical characteristics:

[0067] The main components and calorific value of syngas can be obtained online, but the latent components are difficult to measure stably; after syngas is blended, emission anomalies show sudden or baseline drift characteristics and are difficult to explain by the main components; boiler-side operating data and emission data can form a time-aligned multi-source sequence; emission anomalies have a lag and path dependence on the blending ratio and air distribution adjustment.

[0068] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0069] Figure 1 The application scenarios shown include a biomass gasification device 100, a biomass gas pipeline 101, a coal-fired boiler 102, a flue gas duct 103, an emission monitoring device 104, and a production management system 105, wherein:

[0070] The biomass gasification device 100 is used to generate and output biomass gas, which is transported to the coal-fired boiler 102 via the biomass gas pipeline 101 and participates in combustion together with pulverized coal. The flue gas from the coal-fired boiler 102 is discharged through the flue gas pipeline 103, and an emission monitoring device 104 is installed on the flue gas pipeline 103 to acquire corresponding emission monitoring data. The production management system 105 establishes data interaction connections with the biomass gasification device 100, the coal-fired boiler 102, and the emission monitoring device 104 respectively, to acquire operating data and adjust the operating conditions of the coal-fired boiler 102.

[0071] It is understood that the method of this application can be applied to the production management system 105. Specifically, it can be stored in a storage medium in the form of a computer program. The storage medium includes, but is not limited to, magnetic storage medium, optical storage medium, semiconductor storage medium, and combinations thereof. When the instructions stored in the storage medium are executed by the processor, the production management system 105 executes the parameter optimization method for biomass gas coupled coal-fired boiler based on digital twins as described in this application.

[0072] refer to Figure 2 , Figure 2 A schematic diagram of a processor module provided in an embodiment of this application.

[0073] The processor 1051 is located in the production management system 105. In a specific embodiment not shown in the figure, the processor 1051 can call instructions from the storage medium, execute the corresponding method, and output control instructions, which are then transmitted to the controller to adjust the operating conditions of the biomass gas coupled coal-fired boiler.

[0074] Figure 2 The processor 1051 is shown to include a data acquisition module 10511, a model storage module 10512, an emission prediction module 10513, and an optimization control module 10514, wherein:

[0075] The data acquisition module 10511 is used to acquire online sensing data related to biomass gas and boiler operation data, and to perform time synchronization, validity verification and necessary preprocessing on the data to form input data for subsequent processing.

[0076] The model storage module 10512 is used to store digital twin models, deviation correction models and parameter information related to the operation of biomass gas coupled coal-fired boilers, as well as emission prediction and optimization control, and to provide model calling support for the emission prediction module 10513 and the optimization control module 10514.

[0077] The emission prediction module 10513, based on the input data acquired by the data acquisition module 10511 and the model information stored in the model storage module 10512, predicts the emission change trend within a preset time domain and generates corresponding emission prediction results and constraint boundary information.

[0078] The optimization control module 10514 is used to optimize the control variables based on the emission prediction results output by the emission prediction module 10513, combined with the combustion safety boundary and operating constraints, and generate control commands for adjusting the biomass gas co-firing ratio, total air volume and staged air distribution.

[0079] Next, with reference to the accompanying drawings, the parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins provided in the embodiments of this application will be further described. Figure 3 The method shown is applied to a production management system, which is equipped with a digital twin model. The digital twin model includes a mechanistic sub-model and a deviation correction sub-model. The mechanistic sub-model is used to characterize the mass conservation, energy conservation, and pollutant generation mechanisms of the coupled combustion process of biomass gas and coal-fired boilers. The method includes:

[0080] S1: Acquire sensing data of biomass gas and boiler operation data, and generate a state vector;

[0081] In this embodiment, the biomass gas sensing data includes at least online data that can characterize the energy input characteristics of gaseous fuel and the changing trends of the main components, and the boiler operation data includes at least operation data that can reflect the load level, air distribution status, and oxygen content characteristics in the furnace.

[0082] Those skilled in the art will understand that the sensing data and operational data can be configured according to actual engineering conditions, and their specific types and quantities are not limited, as long as they meet the minimum requirements for characterizing the coupled combustion conditions.

[0083] After time alignment, validity screening, and necessary preprocessing, the data is combined to generate a state vector that characterizes the current coupled combustion conditions. This avoids the information loss problem caused by insufficient characterization by a single parameter and provides a unified state input basis for subsequent model mapping and emission prediction.

[0084] S2: Map the state vector to the simulation results of the digital twin model to obtain the deviation distribution of state variables and emissions under the corresponding operating conditions;

[0085] In this embodiment, the state vector is input as a boundary condition into the digital twin model. The simulation output under ideal interpretable conditions is obtained through the mechanism sub-model. The simulation output is then matched with the measured state quantities and emissions at the corresponding time, thereby forming a sequence of deviation values ​​that reflects the difference between the model output and the actual operation.

[0086] Furthermore, by statistically processing the deviation value sequence, a deviation distribution is constructed to characterize the deviation amplitude and fluctuation characteristics under the current operating conditions. In this way, the influence of implicit components in biomass gas that are difficult to measure directly is indirectly mapped into a deviation structure between multidimensional observables, thereby avoiding the mechanism mismatch problem caused by relying solely on calorific value or principal component equivalence.

[0087] S3: Update the deviation correction sub-model according to the deviation distribution to generate emission predictions within a preset time domain;

[0088] In this embodiment, the deviation distribution is used to update the deviation correction sub-model online, so that emission prediction no longer relies solely on the ideal assumptions of the mechanistic model, but can reflect the impact of actual operating deviations under the current operating conditions.

[0089] Specifically, by incorporating the deviation distribution into the emission prediction process, the emission change trend within the preset prediction time domain is corrected, thereby enabling the early reflection of the direction of emission risk changes even under conditions of latent component fluctuations or sudden changes in operating conditions.

[0090] Those skilled in the art will understand that the length of the prediction time domain can be set according to the adjustment period and control requirements, and this application does not limit it.

[0091] S4: Based on the emission prediction, construct a constrained optimization problem including emission limits and combustion safety boundaries and solve it to obtain the optimized control quantity, wherein the optimized control quantity includes at least one of the following: blending ratio, total air volume, and staged air distribution.

[0092] In this embodiment, based on emission prediction results, emission limits are introduced as constraints into the optimization problem. Combined with operational safety boundaries such as furnace outlet oxygen content, furnace negative pressure, and steam parameters, an optimization solution model under multiple constraints is constructed. By jointly optimizing control variables such as blending ratio, total air volume, and staged air distribution, optimized control quantities are obtained while satisfying emission and combustion safety constraints.

[0093] It should be noted that the digital twin model in this application, as well as the mechanism sub-model and bias correction sub-model configured in the digital twin model, can all be implemented through a combination of model parameterization construction and online update mechanism. Specific configuration methods include, but are not limited to:

[0094] Based on boiler design parameters, fuel characteristic parameters, and operating boundary conditions, a mechanistic calculation model is established to describe the coupled combustion process of biomass gas and coal. Key parameters are initially tuned using offline calibration or historical operating data. Building upon the mechanistic model, a deviation correction structure is introduced to characterize the relationship between the model output and actual operating deviations. This deviation correction structure is trained or fitted using historical operating data to form an initial deviation correction sub-model. The deviation correction structure can be understood as a mathematical mapping or computational logic unit used to characterize the systematic deviation between the mechanistic model output and the actual operating results. Its essence is a parameterized or functionalized correction relationship used to compensate and adjust the mechanistic calculation results; it does not represent a real physical device, control component, or hardware entity structure.

[0095] The deviation correction structure can be implemented in the form of regression function, incremental mapping rule, state-related correction term or equivalent compensation model, etc. Its input is the mechanism model output, operating state characteristics or a combination thereof, and the output is the corresponding deviation estimate or correction amount, which is used to reflect the stable deviation characteristics between the mechanism model and the actual operation in a specific operating condition area.

[0096] Based on the above understanding, the deviation correction structure can be regarded as a functional calculation module added to the mechanistic model to enhance the model's adaptability and engineering practicality under complex operating conditions, fuel quality fluctuations, or equipment aging, without constituting a limitation on the physical system structure.

[0097] In actual operation, the boundary conditions of the mechanism sub-model are updated in real time based on the online sensing data and boiler operation data, and the parameters of the deviation correction sub-model are corrected online or updated on a rolling basis in combination with the deviation distribution.

[0098] Those skilled in the art will understand that the model’s construction method, parameter update frequency, and correction strategy can be flexibly set according to the boiler capacity, fuel source, and operation and management requirements, as long as the dynamic characterization of the coupled combustion conditions and emission response characteristics can be achieved to a minimum. This application does not impose further limitations on this.

[0099] Before detailing the specific technical aspects of the steps, this application's embodiments need to reiterate:

[0100] During the long-term operation of biomass gas coupled with coal-fired boilers, even under conditions of stable fuel supply and no significant fluctuations in online sensing data of main components and calorific value, emission behavior may still exhibit characteristics inconsistent with existing empirical models. These changes do not necessarily manifest as instantaneous exceedances, but rather as slow shifts in the emission response path or increased anomalous sensitivity within local intervals. The fundamental reason lies in the fact that the coupled combustion process itself is a strongly coupled system involving multiple physics fields, with complex interrelationships between fuel input, air distribution, and the in-furnace reaction environment. In such systems, the stability of a single variable cannot guarantee the stability of the overall reaction path; emission behavior is more susceptible to the influence of multi-dimensional state-coordinated changes.

[0101] In engineering practice, traditional control strategies typically rely on closed-loop regulation of a single or a few key indicators, assuming that other unmodeled influencing factors can be treated as random disturbances or covered by empirical margins. However, in scenarios involving biomass gas combustion, due to the unavoidable volatility of fuel sources and the gasification process itself, these unmodeled influencing factors often do not exist independently but act on the combustion process in groups, forming intrinsically correlated change patterns among multiple observables. Simply categorizing them as noise or occasional deviations not only fails to reflect the evolving trends of emission risks in a timely manner but may also introduce unnecessary conservatism at the control level, affecting the economy and continuity of coupled operation.

[0102] This embodiment does not attempt to explicitly model or identify every potential influencing factor in its methodology. Instead, it focuses on the overall operational behavior, shifting the attention to the structural characteristics of the deviation between observable state quantities and emissions. By characterizing the coordinated changes of multidimensional deviations in time and space, it can indirectly reflect whether the current operational state is still within the range that existing models can reasonably explain, without adding additional perceptual burden. This application emphasizes the interrelationships between deviations, rather than the absolute magnitude of a single deviation, thereby avoiding over-adjustment triggered by occasional fluctuations in individual indicators.

[0103] It is important to note that when analyzing the deviation structure, this embodiment does not assume a fixed number or specific physical meaning of the deviation sources. Instead, it maps multidimensional deviations to several representative change patterns, enabling recurring deviation behaviors during operation to be identified and tracked in a unified manner. When a deviation can be stably reconstructed by a finite number of change patterns, the current operating state can be considered to have high interpretability. Conversely, when a deviation exhibits characteristics of being difficult to reconstruct or a significant increase in residual components, it suggests that the operating state may have deviated from the existing empirical range, requiring the introduction of more conservative constraints at the control level.

[0104] It is understandable that the method used in this embodiment does not aim solely at improving the prediction accuracy of a specific emission indicator. Instead, it introduces an intermediate representation that reflects the interpretability of the operating state by characterizing the deviation behavior structure, thus enabling the generation process of emission constraint boundaries to have adaptive characteristics. This processing logic can effectively reduce the impact of sudden emission changes on control stability in application environments with fluctuating biomass gas quality, frequent changes in coupling ratios, or numerous switching of operating conditions, providing a more reliable constraint basis for subsequent optimization solutions.

[0105] Next, we will further elaborate on the technical aspects of the state vector in this application.

[0106] In one example, the sensing data includes the flow rate, temperature, pressure, and lower heating value of biomass gas, as well as the concentrations of the main components corresponding to carbon monoxide, hydrogen, methane, and carbon dioxide. The boiler operating data includes unit load, coal feed rate, primary air volume, secondary air volume, total air volume, furnace outlet oxygen content, furnace negative pressure, main steam flow rate, main steam pressure, main steam temperature, and reheat steam temperature.

[0107] It is understood that the sensed data can be collected by online detection devices installed at biomass gas transmission pipelines or relevant process nodes. These online detection devices include, but are not limited to, flow meters, temperature sensors, pressure sensors, calorific value analyzers, and gas component analyzers. Any of these sensed data points can be directly measured or obtained through online analysis under existing engineering conditions. For example, the lower heating value can be obtained through component analysis results or dedicated calorific value analysis equipment, and the concentration of the main component can be measured through online analysis methods such as infrared, thermal conductivity, or chromatography.

[0108] Those skilled in the art will understand that the above-described sensing method is merely an example. As long as data reflecting the energy characteristics and compositional trends of biomass gas can be continuously acquired during operation, it can meet the implementation requirements of the method in this application, and this application does not limit it in this regard.

[0109] In addition, the boiler operating data can also be obtained through the power plant's existing monitoring and control units, such as distributed control devices or equivalent operation monitoring platforms, for real-time recording and management of the unit's operating status. All of the above-mentioned boiler operating data are process variables naturally generated during boiler operation, and their acquisition does not depend on additional dedicated detection equipment.

[0110] Those skilled in the art will understand that the specific types, sampling frequencies, and accuracy of the boiler operating data can be configured according to the unit capacity, operating mode, and management requirements. It is sufficient that the data can reflect the changes in boiler load level, air distribution status, and combustion environment in the furnace to a minimum. This application does not impose any further limitations on this.

[0111] In another example, the process of generating state vectors is not a simple data concatenation, but rather a construction centered around the characterizability and computability of coupled combustion conditions.

[0112] In actual operation, biomass gas sensing data and boiler operation data originate from different measurement stages, resulting in objective differences in their sampling periods, timestamp accuracy, and signal stability. Direct joint analysis can easily introduce spurious correlations caused by time misalignment or instantaneous anomalies. Therefore, in this embodiment, the biomass gas sensing data and boiler operation data are first aligned to a unified time reference through interpolation, resampling, or window alignment to achieve time synchronization. Simultaneously, outliers that significantly deviate from physically reasonable ranges are removed or corrected based on engineering experience and statistical characteristics. Furthermore, different physical quantities are normalized or standardized for dimensional uniformity. These processes ensure comparability of various data types on the same time and numerical scales, laying the foundation for subsequent feature extraction.

[0113] Furthermore, after completing the data preprocessing, this embodiment further introduces principal component analysis to jointly process the processed sensing data and boiler operation data.

[0114] It should be noted that in biomass gas coupled combustion scenarios, there are a large number of original variables, and these variables are generally correlated. For example, fuel input, air distribution, and oxygen distribution in the furnace often exhibit a synergistic relationship. Directly using all original variables as model input would not only increase computational complexity but may also weaken the model's sensitivity to changes in key operating conditions. Therefore, principal component analysis is used to reduce the dimensionality of the multidimensional data, mapping the original variables into a small number of comprehensive features while preserving the main information. These features statistically reflect the combustion boundary conditions, oxygen distribution characteristics in the furnace, and heat input trends, thus avoiding interference from irrelevant or redundant information in subsequent modeling and prediction.

[0115] Those skilled in the art will understand that the specific dimension to be retained in principal component analysis can be selected based on the cumulative contribution rate or engineering requirements, and this application does not impose strict limitations on this.

[0116] Furthermore, after obtaining the aforementioned characteristic quantities, this embodiment combines them in chronological order to form a state vector characterizing the current coupled combustion condition. This state vector is not a static parameter set, but rather a dynamically updated condition characterization quantity that changes with the operating state. It contains at least two types of information: one type reflects the energy input and main component variation characteristics of biomass gas, used to characterize the influence of the fuel side on the combustion process; the other type reflects the boiler load level, air distribution status, and in-furnace reaction environment, used to characterize the constraints imposed on combustion behavior by the equipment and operating sides. The state vector constructed in this way can fully characterize the key features of the coupled combustion condition while maintaining dimensional control, providing a stable and physically meaningful input basis for subsequent digital twin model simulation mapping, deviation analysis, and emission prediction. This enables those skilled in the art to implement corresponding engineering applications based on the technical content of this embodiment.

[0117] Next, we will further elaborate on the technical aspects of the simulation results of the method in this application.

[0118] It is understood that the simulation results in this application are derived from the online simulation output of the digital twin model. The basic principle is to abstract the physical mechanism of the coupled combustion object into a computable mathematical description, and combine it with the real-time updated operating boundary conditions to synchronously map and deduce the actual operating state.

[0119] Specifically, the digital twin model is not a static replica of the boiler system. Instead, it uses a mechanistic sub-model as its core, continuously receiving current operating condition information represented by state vectors. It takes factors such as the energy input from biomass gas and coal combustion, air distribution, and load levels as boundary conditions. By solving the mechanistic relationships related to mass conservation, energy conservation, and pollutant generation, it obtains the theoretical response results for each state quantity and emission under the current operating condition assumptions. The simulation results obtained represent the operational behavior that the coupled combustion process should exhibit under given mechanistic assumptions and ideal interpretability conditions.

[0120] It should be noted that in actual engineering operations, due to fluctuations in fuel quality, equipment aging, and the existence of hidden influencing factors, deviations inevitably exist between actual operating behavior and pure mechanistic calculation results. This application does not attempt to completely eliminate such deviations by continuously complicating the mechanistic model, but rather uses the simulation results as a benchmark response to reflect the interpretable operating trends of the system under current operating conditions. By comparing the simulation results with measured state variables and emissions, the structural differences between the mechanistic model and actual operation can be intuitively characterized, providing a basis for subsequent deviation analysis and correction. This approach retains the advantages of the mechanistic model in terms of extrapolation and physical consistency, while avoiding excessive reliance on complex or difficult-to-obtain online factors.

[0121] In this embodiment, the simulation results are not generated once and used permanently, but are generated continuously during operation as the state vector is updated. That is, when the biomass gas composition, calorific value, or boiler operating state changes, the corresponding boundary conditions will be applied synchronously to the digital twin model, causing it to output new simulation results.

[0122] In one example, the process of generating the above simulation results further incorporates historical operational information to enhance the adaptability and stability of the digital twin model under complex operating conditions.

[0123] Specifically, after obtaining the state vector characterizing the current operating state, the system first extracts boundary conditions reflecting key constraints on both the fuel and equipment sides, such as the energy input level of biomass gas, the composition characteristics of the main components, and the unit load range. Based on this, the boundary conditions are applied as input to the digital twin model, and initial simulation results corresponding to the current operating condition are calculated through the mechanistic sub-model. These initial simulation results primarily reflect the theoretical response trend of the system under these boundary conditions under given mechanistic assumptions, serving to provide a physically consistent reference starting point for subsequent deviation corrections.

[0124] Based on this, this embodiment does not directly use the initial simulation results as the final output, but further introduces historical operating information similar to the current operating conditions. Within a preset time window, historical operating condition data that meets the similarity of the operating condition boundary conditions is selected from the historical operating records. The historical operating condition data includes at least the boiler operating data and actual emission data for the corresponding time period.

[0125] It should be noted that the length of the time window and the selection criteria for historical operating conditions can be set according to operational stability and data accumulation, as long as they can cover representative operating segments. By comparing the initial simulation results with the actual operating results under the historical operating conditions, a reference deviation reflecting the degree of systematic deviation of the mechanism model in this type of operating condition region can be obtained, thereby avoiding the random influence caused by relying solely on data from a single moment to correct the model.

[0126] Furthermore, when correcting the initial simulation results, this embodiment does not simply superimpose historical deviations with equal weights, but introduces a difference-aware weighted fusion mechanism based on the degree of difference between the current target working condition boundary conditions and the boundary conditions of each historical working condition.

[0127] Specifically, by comparing key boundary parameters such as the lower heating value of biomass gas, the concentration of main components, and unit load, the similarity between the current operating condition and historical operating conditions is quantified, and corresponding weighting coefficients are determined accordingly. Historical operating conditions with higher similarity have higher weights in the correction process, while the impact of historical operating conditions with greater differences on the correction results is weakened. The corrected simulation results obtained in this way inherit the advantages of the mechanistic model in terms of physical consistency and fully absorb the actual response characteristics verified in historical operation. This allows the digital twin model to output simulation results that are closer to actual operation even under operating condition switching or fuel quality changes, thus providing a more reliable foundation for subsequent deviation analysis, emission prediction, and optimized control.

[0128] For example, the following numerical example is provided to illustrate the calculation process of obtaining the weighting coefficients. This example is only used to explain the calculation chain, field organization, and dimensional relationships. The selected parameters and values ​​are illustrative and do not represent actual calibration results or engineering recommendations.

[0129] In one example, the key parameters used to characterize the boundary conditions of the target working condition are organized into a boundary condition vector. The contribution weight of historical operating conditions to the reference deviation is determined using a link of "difference degree - similarity degree - weighting coefficient". For ease of explanation, the boundary condition vector uses parameters such as lower heating value, principal component concentration, and unit load, defined as: ,in The lower heating value of biomass gas, For example, the volume fraction of the corresponding component. This is the volume fraction of carbon monoxide. The integral number of hydrogen gas. This represents the volume fraction of methane. This represents the volume fraction of carbon dioxide. To eliminate the influence of dimensions for the unit load, scale normalization is performed on each dimension before calculating the degree of difference. For example, minimum-maximum normalization is performed using an engineering feasible interval. The given interval for this example is as follows: , , , , , .

[0130] In this example, the target working condition boundary conditions Set as: Three historically similar operating conditions are selected within a preset window, with the following boundary conditions:

[0131] ;

[0132] ;

[0133] ;

[0134] After normalization, Euclidean distance is used as a measure of dissimilarity. The measure of dissimilarity can be understood as the comprehensive deviation from multidimensional boundary conditions. The calculated dissimilarity... , ,as follows: , , The degree of difference is shown in the example calculation results, which are easy to understand. Indicates correspondence The degree of difference is further mapped to similarity, and a weighted coefficient is formed accordingly. For example, similarity is defined using an exponential decay form. :

[0135] ;

[0136] in, This indicates that the smoothing factor is used to adjust the sensitivity of the "distance difference" to the weights. In this example, it is set to 0.25. , , And by normalizing the similarity, a weighted coefficient is obtained, which is easy to understand. Indicates correspondence The similarity shows that historical conditions with boundary conditions closer to the target condition have a higher weight in the correction process, while the weight of historical conditions with greater differences is suppressed.

[0137] Next, we will further elaborate on the technical content of the method of this application regarding the deviation distribution.

[0138] In one example, the state vector is mapped to the simulation results of the digital twin model to obtain the deviation distribution between the state variables and emissions under the corresponding operating conditions, including:

[0139] S2.1: Determine the set of mapping variables corresponding to the state vector, wherein the mapping variables include derived state variables used to characterize the coupling relationship between fuel and wind distribution and check emissions used to characterize the emission response;

[0140] Specifically, the state vector describes the current coupled combustion condition, but the variables used for mapping and verification are not the same as the state vector. This is because the state vector typically contains selected or dimensionally reduced features, focusing on the overall characterization of the condition; while the parameter mapping process requires selecting variables with engineering observability, alignability, and sensitivity to the fuel-air coupling relationship, enabling simulation and experimental results to be compared at the same semantic level. This avoids situations where features cannot be directly observed, leading to a lack of verification, or where overly narrow variable selection masks the source of bias. Therefore, the mapping variable set uses a combination of derived state variables and verification emissions, ensuring that the mapping covers both the changes in the internal state of the combustion process and the resulting response at the emission end.

[0141] It should be noted that the mapping variable set adopts a combination of derived state variables and verification emissions. The specific combination logic is as follows: first, a derived state quantum set on the process side is constructed based on the fuel-air distribution coupling mechanism; then, a verification emission quantum set on the result side is constructed based on the constrainable monitoring quantities at the emission end. The two are merged into the same mapping set according to the principle of "process alignment first, result constraint second". The derived state variables are used to cover the observable characterization of key links in the combustion process. Variables that can reflect the equivalent energy input of fuel, the overall oxygen surplus, the staged air distribution organization mode, and the thermal load / heat balance drift are selected first, so that the simulation conditions and the measured conditions are consistent in engineering semantics. The verification emissions are used to cover the explicit response. Nitrogen oxides and carbon monoxide, which are sensitive to burnout and air distribution, are included first. When conditions permit, sulfur dioxide and particulate matter are introduced as lateral constraints to distinguish the different contributions of changes in fuel input, end-of-pipe treatment conditions, and combustion organization to emissions, thereby avoiding the masking of process-side bias sources by fitting only a single emission index. In its implementation, the mapping variable set is not simply stacked, but rather established according to engineering meanings, with consistent alignment standards: derived state variables and emissions are synchronously aggregated within the same time period, considering the response lag of the emission chain when necessary; quantities that fluctuate significantly with load are converted or normalized using a consistent standard to eliminate apparent biases caused by dimensional differences and load disturbances; during mapping and verification evaluation, process-side derived state variables are given higher alignment priority to lock in a reasonable range of fuel-air distribution coupling states, and emissions are then constrained within this range to filter out candidate solutions that do not meet emission response consistency. Through this combination method, the mapping set can constrain both the correctness of internal state evolution and the acceptability of emission-side response results, achieving synchronous verification of "mechanism consistency" and "response consistency".

[0142] In this embodiment, the derived state variables are verifiable variables obtained from collected boiler operation data and biomass gas sensing data using an engineering-consistent calculation method, used to characterize the coupling relationship between fuel input and air distribution organization. The derived state variables may include, but are not limited to:

[0143] The equivalent heat input, calculated from the coal feed rate, biomass gas flow rate, and lower heating value, is used to reflect changes in the total energy supply on the fuel side. The equivalent excess air coefficient or equivalent air-fuel ratio, calculated from the total air volume and equivalent heat input, is used to reflect the degree of surplus of air distribution relative to heat input. The primary air ratio, secondary air ratio, or staged air distribution ratio, calculated from the primary air volume, secondary air volume, and total air volume, are used to reflect the staged combustion organization mode. The oxygen margin characterization quantity formed by the combined oxygen content at the furnace outlet and the total air volume is used to reflect the oxygen-rich or oxygen-deficient trend of the combustion environment. The thermal load characterization quantity constructed from the main steam flow rate, main steam pressure, main steam temperature, and reheat steam temperature is used to reflect whether the boiler thermal balance state has drifted.

[0144] The verification emissions are emission response quantities obtained from emission monitoring and that can be used for constraints. They may include one or more of nitrogen oxides, carbon monoxide, sulfur dioxide, and particulate matter. In cases where there is a carbon monoxide measuring point at the furnace outlet, carbon monoxide on the furnace side may also be included as verification emissions in order to reflect the trend of incomplete combustion in advance.

[0145] S2.2: Obtain the corresponding measured values ​​for each mapping variable in the set of mapping variables, and match the measured values ​​with the simulation results output by the digital twin model at the same time.

[0146] Specifically, the comparability of measured and simulated values ​​depends on the consistency of the time base and the semantic consistency of the measurement points. Operational data and emission data often have different sampling periods, and some variables have physical lags. For example, emission monitoring values ​​have transmission and analysis delays relative to the in-furnace reaction. Without matching correction, the deviation sequence will be mixed with spurious fluctuations caused by sampling misalignment, which will amplify or distort the deviation distribution.

[0147] In this embodiment, data acquisition and alignment criteria are established for each mapping variable. For boiler operation data variables, a resampling method based on the control cycle is adopted to mean or medianize high-frequency data according to the control cycle, avoiding single-point spike interference. For low-frequency data, interpolation or hold strategies are used to achieve alignment, ensuring stable values ​​over the control cycle. For emission monitoring quantities, fixed or adaptive time delay compensation is configured by combining flue gas passage length, sampling probe position, and instrument processing time to back-align emission measurements to the corresponding furnace operating time. When precise delay information is lacking, a correlation-based delay estimation method is used, finding the time offset with the highest correlation between emissions and furnace outlet oxygen, total air volume, or carbon monoxide within the historical window as the alignment offset. For simulation output, the simulation values ​​corresponding to the mapping variables are output at a time step consistent with the control cycle. When the simulation output and the measured time are inconsistent, linear interpolation or adjacent-step hold methods are used to pair them under the same time index.

[0148] S2.3: Calculate the difference between the matched measured values ​​and simulation results to obtain the deviation values ​​corresponding to each mapping variable;

[0149] Specifically, deviation values ​​cannot be simply subtracted and used directly for subsequent statistics. This is because the dimensions, numerical ranges, and allowable fluctuation ranges of different mapping variables vary significantly, and some variables have symbolic meanings. For example, the positive or negative sign of furnace negative pressure deviation reflects changes in the direction of suction strength. If the deviation calculation method is not standardized, the deviation distribution will be dominated by dimensions, causing the statistical results to lose their ability to express the relative degree of deviation. Furthermore, the deviations of emissions and state quantities have different engineering meanings.

[0150] Emissions deviations directly correlate with compliance risks, while state deviations are more often used to explain combustion organization shifts. Deviation definitions need to balance interpretability with operability for subsequent margin generation.

[0151] In this embodiment, a categorized difference rule is used for different types of mapping variables. For variables with definite physical units and stable dimensions, such as flow rate, temperature, pressure, main steam parameters, and furnace outlet oxygen content, the absolute deviation between the measured value and the simulated value is used as the deviation value, and the sign of the deviation is retained to reflect the direction of deviation. For emission verification quantities, such as nitrogen oxides, carbon monoxide, sulfur dioxide, and particulate matter, in addition to the absolute deviation, a relative deviation sequence can be further formed. That is, the deviation is proportionalized with the simulated value or limit value as a reference to maintain the comparability of deviations in different load ranges. When the simulated value is close to zero or in a low value range, a proportionalization method with the limit value or historical baseline as a reference is used to avoid meaningless amplification of the proportional deviation.

[0152] S2.4: Summarize the deviation values ​​according to preset statistical rules to generate a deviation distribution that characterizes the deviation amplitude and fluctuation range under the target working condition, wherein the statistical rules include at least one of sliding time window statistics and quantile statistics;

[0153] Specifically, the deviation distribution is used to characterize the amplitude range and fluctuation characteristics of deviations under target operating conditions. Its construction method needs to balance sensitivity to short-term fluctuations and robustness to outliers. If only single-point deviations or short-term averages are used as the characterization, they are easily dominated by random noise or transient disturbances; if only full-time statistics are used, the phased changes brought about by operating condition switching, load ramping, or control actions are easily mixed into the same distribution, resulting in a lack of interpretability in the distribution width. Therefore, the deviation distribution adopts a combination of sliding time window statistics and quantile statistics to ensure a close fit to the current operating conditions while tolerating outliers.

[0154] In this embodiment, the sliding time window is updated on a rolling basis according to the control cycle. The window length is set to a combination of several control cycles based on the unit's dynamic response characteristics and emission monitoring lag, enabling the window to cover the main dynamic process between a single adjustment action and the emission response. For the deviation value sequence of each mapped variable, the central tendency and dispersion of the deviation are calculated within the window. The central tendency can be represented by the mean or median, and the dispersion can be represented by robust statistics such as interquartile range, range, or mean absolute deviation. For emission-related verification quantities, the upper quantile is preferentially used as the upper bound of fluctuation to match the risk preference of the constraint boundary. Quantile statistics are used to give the upper and lower bounds of the deviation within the window, for example, using the higher quantile as the upper bound of the deviation and the lower quantile as the lower bound of the deviation, and simultaneously recording the maximum and minimum values ​​of the deviation within the window for fault diagnosis or abnormal alarm purposes. For windows with obvious operating condition switching, the window can be divided into steady-state segment and transition segment according to the load change rate or air distribution change rate, and statistically analyzed separately and then weighted and synthesized according to the time proportion to avoid unnecessary expansion of the steady-state distribution by the transition segment.

[0155] Next, we will further elaborate on the technical aspects of the emission prediction method in this application.

[0156] It should be noted that the deviation correction sub-model described in this application is not a simple correction step independent of the simulation model, but rather an intermediate layer connecting the deviation distribution and emission prediction results, used to transform the deviation characteristics identified under the current operating conditions into a constraint description of future emission behavior.

[0157] Therefore, the emission constraint boundary can be understood as:

[0158] Under given operating conditions and deviation statistical characteristics, the dynamic limit range formed by the upper limit level that each check emission may reach within the preset prediction time domain is not equivalent to a fixed emission limit, but a variable boundary that is updated with changes in operating conditions and deviation structure.

[0159] Specifically, in engineering operations, emission limits are typically given as operational management requirements, and their values ​​remain constant over a long timescale. However, the sensitivity of emission response to operational adjustments varies with fuel characteristics, air distribution, and load levels. If fixed limits are used as the sole constraint during the emission prediction phase, the prediction results often lack lead time when deviation fluctuations intensify or hidden influencing factors increase, and risks are easily reflected only when the limits are approached. This application introduces emission constraint boundaries based on deviation distribution, enabling the prediction algorithm to simultaneously consider the dispersion and uncertainty range of the emission response under current operating conditions when calculating future emission trends, thereby reflecting potential risks before the limits are reached.

[0160] In one example, emission predictions are generated using a prediction algorithm, specifically including:

[0161] S3.1: Based on the deviation distribution, determine the fluctuation range of each verification emission within the preset time domain, and convert the fluctuation range into a safety margin;

[0162] Specifically, the deviation distribution is used to express the shape and degree of dispersion of the deviation between the digital twin simulation output and the measured emissions near the target operating condition. Since emission signals are typically affected by factors such as load fluctuations, air distribution regulation, fuel input changes, and measurement lag, single-point deviations or instantaneous extreme values ​​lack stable representativeness. Directly using them for subsequent constraint construction can easily introduce overly conservative or overlooked risks. Therefore, when determining the fluctuation range of each check emission within a preset time domain, interval representation is used based on the statistical characteristics of the deviation distribution. This ensures that the fluctuation range simultaneously covers deviations that may occur under normal fluctuations and operating condition disturbances, and maintains a time scale consistent with the control period and prediction time domain.

[0163] In this embodiment, the corresponding deviation distribution parameters are read for each verification emission. These deviation distribution parameters may include the upper quantile deviation, lower quantile deviation, median deviation, and statistics used to characterize the degree of dispersion within the sliding time window. The process of determining the fluctuation range adopts a segmented approach: when the target operating condition is in a relatively stable range, the upper and lower quantile deviations within the sliding window together constitute the fluctuation range to ensure that the range covers normal fluctuations; when the target operating condition is in a transitional range such as load ramping, co-firing ratio adjustment, or air distribution reconfiguration, a rate of change constraint is introduced based on the above quantile range, and the fluctuation range is expanded by combining the change amplitude of the deviation distribution of adjacent windows, so that the fluctuation range can include short-term deviations caused by operating condition switching. Subsequently, when converting the fluctuation range into a safety margin, a fixed ratio conversion is not used. Instead, the conversion caliber is determined based on the emission dimensions and control significance: for constrained emissions such as nitrogen oxides and sulfur dioxide, the upper limit of the deviation fluctuation range is used as the main source of safety margin; for emissions such as carbon monoxide, which are more closely related to combustion stability, in addition to considering the upper limit, the safety margin is also compensated by combining the width of the fluctuation range, so as to avoid insufficient constraints caused by still using a single upper limit to characterize the fluctuation when it intensifies.

[0164] S3.2: Combine the safety margin with the preset emission limit to obtain the emission limit that varies with the target operating condition, wherein the emission limit is less than or equal to the preset emission limit;

[0165] Specifically, emission limits are typically fixed thresholds, while the fluctuation of emission response under coupled combustion conditions varies with operating conditions. If fixed limits are used directly as the constraint boundary, prediction and control often only manifest risks when approaching the limits, resulting in a lack of lead time for adjustment. By combining safety margins with emission limits to form an upper limit for emissions, the constraint boundary can be adjusted according to changes in target operating conditions. This provides a buffer for adjustment when emission response fluctuations increase, and avoids over-tightening when fluctuations are small.

[0166] In this embodiment, the emission upper limit is generated by a limit deduction method. That is, at each prediction step, the corresponding safety margin is deducted from the emission limit to obtain the emission upper limit. When the safety margin is too large, which may lead to an excessively low emission upper limit and thus affect the continuity of coupled operation, a lower limit protection mechanism is introduced to set a minimum allowable value for the emission upper limit. The minimum allowable value can be obtained from the operation management strategy or historical compliance statistics to avoid the constraint boundary being tightened to an unachievable level under abnormal data. For different verification emissions, the limit deduction caliber can be set differently: for emissions where compliance is the primary constraint, the safety margin deduction adopts a margin consistent with the upper bound of the deviation; for emissions strongly correlated with combustion stability, a comprehensive margin weighted by the upper bound margin and the fluctuation width can be used, so that the emission upper limit reflects both the risk of exceeding the limit and the control margin requirement when the combustion state is unstable.

[0167] S3.3: Using the emission upper limit as the emission constraint boundary, an emission prediction within the preset time domain is generated by combining the emission constraint boundary with a prediction algorithm;

[0168] Specifically, emission prediction not only needs to provide future emission trends, but also needs to reflect the risk-limiting effect of constraint boundaries during the prediction process. If the prediction algorithm only extrapolates from historical sequences without introducing emission constraint boundaries, the prediction results may drift unreasonably near the boundaries, making it difficult for subsequent optimization solutions to form a stable and usable feasible region. If the prediction results are only truncated after the boundary, the continuity of the prediction curve will be disrupted, making the gradient information or rolling search process of the optimization solution unstable.

[0169] In this embodiment, the prediction algorithm uses the current state vector, digital twin simulation output, and deviation correction information as inputs to gradually generate emission prediction sequences for future time periods according to a preset prediction time domain. During the prediction process, for each prediction step, the corresponding emission upper limit is read as a boundary constraint, and its compliance with boundary requirements is simultaneously verified when generating the predicted value for that step. When the predicted value approaches the boundary, a boundary-driven correction strategy is introduced to suppress the prediction increment, ensuring that the prediction curve remains continuous near the boundary and avoids crossing it. The correction strategy can be implemented using increment limiting, boundary distance weighting, or trend decay, ensuring that the prediction result maintains its responsiveness to changes in operating conditions while satisfying the boundary requirements. For emissions with measurement lag, lag compensation is introduced synchronously in the prediction algorithm, applying the boundary constraints to the compensated equivalent emission sequence to avoid prediction deviations caused by boundary alignment errors.

[0170] It is understood that the prediction algorithm in this application is not limited to a specific form, but refers to a calculation method that can extrapolate the emission change trend within a preset time domain under the condition of introducing emission constraint boundaries. The prediction algorithm can be a state-based recursive prediction method, such as a prediction method that uses the current state vector and simulation output to perform rolling updates over time; it can also be a time series prediction method based on historical data, such as extrapolating future emissions by modeling historical emission data and its relationship with operating status, and combining it with current operating condition information; or it can be a hybrid prediction method that combines mechanism calculation results with data-driven prediction, by superimposing correction amounts guided by deviation correction information on the basis of mechanism prediction to achieve dynamic correction of emission trends.

[0171] Those skilled in the art will understand that the specific implementation of the above prediction algorithm can be selected according to the computing resources, prediction time domain length and control accuracy requirements. As long as the emission constraint boundary can be introduced in the prediction process and the emission prediction result corresponding to the prediction time domain can be generated, the implementation requirements of the method of this application can be met, and this application does not make any further limitations in this regard.

[0172] In one example, a commonly used and easily implemented online prediction algorithm is chosen: an autoregressive model with exogenous input. This model is combined with digital twin simulation output and bias correction information to form a hybrid prediction of baseline response and residual recursion. The autoregressive model is characterized by low computational cost and rolling updates, making it easy to synchronously introduce emission constraint boundaries and continuously correct them within each prediction step.

[0173] For example, taking NOx emissions as the prediction target, the preset prediction time domain is the next 5 minutes, the prediction step size is 1 minute, and there are a total of 5 prediction points. The upper limit of this emission is set at 200 mg / Nm³.

[0174] To demonstrate the role of boundary constraints in limiting risk during the prediction process, this example does not employ ex-post truncation. Instead, it performs boundary-driven corrections simultaneously when generating prediction values ​​at each step, ensuring that the prediction curve is continuous and differentiable near the boundary, i.e., continuously usable at least in the sense of discrete step length, and avoiding crossing the upper limit.

[0175] The specific calculation process can be executed according to the following logic:

[0176] First, lag compensation is applied to emissions with measurement lag to obtain the equivalent current emissions for prediction. For example, online NOx has a lag of about 1 minute. The current measured NOx reading is 176 mg / Nm³, but this value corresponds to the actual emissions 1 minute ago. In this case, combining the simulated change rate of the digital twin in the most recent minute and the deviation correction information, the current equivalent NOx is estimated with compensation, resulting in: the simulated NOx in the digital twin increased by approximately +6 mg / Nm³ in the past minute; the systematic deviation correction for this operating condition is approximately +4 mg / Nm³ lower than the simulation, meaning the measured value is 4 mg / Nm³ higher than the simulation.

[0177] Based on this, the current equivalent NOx is approximately 186 mg / Nm³. All subsequent boundary checks will be performed using this equivalent NOx sequence to avoid boundary alignment errors.

[0178] Then, construct the predicted baseline term and residual term, where:

[0179] The baseline terms are derived from the mechanism deduction output of the digital twin under the current state vector and boundary conditions for the next 5 minutes. For example, the simulation baseline NOx for each minute in the future is: 188 after 1 minute, 191 after 2 minutes, 194 after 3 minutes, 197 after 4 minutes, and 199 after 5 minutes. The unit of the simulation baseline is mg / Nm³.

[0180] The residual term is used to characterize the dynamic deviation of actual operation from the mechanistic baseline, and ARX is used to recursively calculate the residual. The exogenous input of ARX can be selected from key driving quantities in the state vector, such as load, primary air / secondary air ratio, O2, gas calorific value or blending ratio, etc., and parameters can be updated online with the most recent window data or using identified parameters.

[0181] For ease of explanation, let's assume that the ARX residuals identified from the most recent operating window exhibit the following recursive behavior: Under conditions of increased load and slightly increased gas calorific value, the residuals show a slow upward trend, which means that the "unconstrained original predicted values" may have the following sequence after being superimposed relative to the baseline term: 193 after 1 minute, 197 after 2 minutes, 202 after 3 minutes, 206 after 4 minutes, and 209 after 5 minutes, all in mg / Nm³.

[0182] As you can see, the predicted value will cross the upper limit of 200 starting from the 3rd minute.

[0183] Next, a boundary-driven correction strategy is introduced to suppress the prediction increment at each step, rather than truncating the predicted value. An exemplary implementation using a boundary distance-weighted trend decay approach is shown below:

[0184] When generating the next predicted value, first calculate the original predicted increment for that step, that is, the increment relative to the predicted value of the previous step;

[0185] Next, calculate the safe distance between the current predicted value and the upper limit, which is the upper limit of 200 minus the current predicted value;

[0186] When the safety distance is greater than the preset buffer zone, for example, 10 mg / Nm³ can be used without suppression; when entering the buffer zone, the next increment is attenuated according to the safety distance ratio; when the original prediction will cause the boundary to be crossed, the increment is compressed to just not cross the boundary and retain a small margin, for example, 0.5 mg / Nm³, so as to ensure that the curve is continuous and does not cross the boundary near the boundary.

[0187] Taking the original predicted sequence above as an example, assuming the buffer band is 10 mg / Nm³, then:

[0188] The prediction for the first minute is 193, which is 7 points away from the upper limit and has entered the buffer zone; however, the point has not yet crossed the boundary and can be retained.

[0189] The original prediction for the second minute was 197, which is 3 away from the upper limit. It continues to enter the buffer zone and is allowed to remain.

[0190] The original prediction for the 3rd minute was 202, which would exceed the limit. Instead of truncating to 200, we reduce the increment from the 2nd to the 3rd minute to ensure the 3rd minute falls below the upper limit while retaining a margin, for example, by revising it to 199.5.

[0191] The original trend is still rising in the 4th minute, and if we directly extrapolate, it will still go out of bounds; since we are now near the boundary, the increment is further weakened, and the prediction can remain around 199.6.

[0192] The same applies to the 5th minute; the prediction remains around 199.7.

[0193] Therefore, an example of a continuous emission prediction sequence that satisfies the boundary constraints is obtained as follows:

[0194] 193, 197, 199.5, 199.6, 199.7, where the unit is mg / Nm³.

[0195] It should be noted that the impact of the boundary-driven correction on the prediction is reflected in the generation process. In other words, each predicted value is formed after reading the emission upper limit and correcting the increment, thereby avoiding two types of engineering problems: first, unreasonable drift near the boundary leading to instability in the feasible region; second, post-hoc truncation causing discontinuity in the curve and oscillations in the rolling optimization solution process. For example, the algorithm corresponding to rolling optimization can be the gradient method or a heuristic search algorithm.

[0196] Furthermore, when operating conditions change during the forecasting process, such as a decrease in load or an adjustment in air distribution leading to a downward trend in NOx, the exogenous inputs and digital twin baselines of the ARX will change synchronously, and the original forecast increment may turn negative. In this case, the boundary-driven correction will not "force the value to conform to the boundary," but will allow the forecast value to fall naturally from near the boundary, thereby maintaining its responsiveness to changes in operating conditions. This allows the forecast to satisfy constraints without sacrificing controllability and accuracy.

[0197] Next, we will further elaborate on the technical content of the method in this application regarding the optimization of control quantities.

[0198] In one example, a constrained optimization problem, including emission limits and combustion safety boundaries, is constructed and solved based on the emission predictions, including:

[0199] S4.1: The weights of the objective function are determined by the analytic hierarchy process to obtain the optimized objective function, wherein the optimized objective function is used to reduce the combustion regulation cost while satisfying the emission constraint boundary.

[0200] S4.2: The emission prediction values ​​at each time point within the preset time domain and the corresponding emission constraint boundaries at each time point constitute emission constraint conditions;

[0201] S4.3: Determine the combustion safety boundary based on the boiler operating data and make the combustion safety boundary a safety constraint condition, wherein the combustion safety boundary includes at least one of the following: lower limit of oxygen content at furnace outlet, allowable range of negative pressure in furnace, upper limit of main steam temperature, and upper limit of reheat steam temperature;

[0202] S4.4: The optimized control quantity is determined as a decision variable and the range and rate of change of the decision variable are limited, wherein the decision variable includes at least one of the following: co-firing ratio, total air volume, and staged air distribution;

[0203] S4.5: In each control cycle, using the optimized control quantity of the previous control cycle as the initial value, and combining the emission constraints and the safety constraints, the decision variables are solved in a rolling manner through the optimization objective function to obtain the optimized control quantity of the current control cycle.

[0204] It should be noted that the above-described constraint optimization solution process is carried out under the premise that emission prediction results have been given and emission constraint boundaries have been determined according to the target operating conditions. Its core idea is not simply to pursue the optimal value of a single control variable, but rather to find a compromise solution for operational adjustments under multiple objectives and constraints, thereby achieving a sustainable balance between emission control, combustion safety, and adjustment costs. In actual operation, adjustments to the co-firing ratio, air distribution method, and total air volume often simultaneously affect emission levels, in-furnace combustion stability, and equipment heat load. If only emission minimization is taken as the objective, it can easily lead to frequent adjustments or deviations from the safe operating range; if only operational stability is taken as the objective, the biomass gas co-firing ratio or emission margin may be sacrificed. Therefore, the importance of different optimization objectives is assessed in a structured manner using the analytic hierarchy process (AHP), transforming factors such as the degree of emission constraint satisfaction, the maintenance of combustion stability, and the adjustment range into a weighted comprehensive objective function, so that the optimization solution can reflect the true trade-offs in operation management.

[0205] In this embodiment, the objective function is constructed primarily to reduce the cost of combustion regulation. This cost includes not only the absolute changes in control quantities such as blending ratio, total air volume, or staged air distribution, but also the impact of the rate of change of these control quantities on the continuity of the combustion process. Using the analytic hierarchy process (AHP), emission risk, operational safety, and regulation stability are decomposed into several hierarchical factors. Based on operational experience or management strategies, each factor is compared in pairs to obtain relative weights, which are then synthesized into the objective function weight vector. The emission prediction results at each time point within a preset time domain, together with the corresponding emission constraint boundaries, constitute the emission constraints, ensuring that the optimization process meets the requirement of not exceeding the constraint boundaries at each prediction step. Simultaneously, the combustion safety boundary determined based on boiler operating data is introduced as a hard constraint into the optimization problem. For example, by limiting the oxygen content at the furnace outlet to a given lower limit, maintaining the furnace negative pressure within an allowable range, and ensuring that the main steam and reheat steam temperatures do not exceed the safety upper limit, the optimization results are guaranteed to always remain within the equipment's permissible operating range.

[0206] Furthermore, in terms of setting decision variables, this embodiment uses directly adjustable quantities such as the co-firing ratio, total air volume, and staged air distribution, which have a significant impact on emissions and combustion status, as optimization variables. A value range and a rate of change constraint are set for each decision variable. The value range reflects equipment capacity and process limitations, while the rate of change constraint prevents excessively rapid adjustments in adjacent control cycles, which could lead to combustion instability or mechanical shocks. Within each control cycle, the optimized control quantity from the previous control cycle is used as the initial value for rolling optimization, ensuring temporal continuity and reducing solution fluctuations. By solving the objective function under the combined effects of emission and combustion safety constraints, the feasible optimized control quantity for the current control cycle is obtained. This rolling optimization method allows the control strategy to be gradually adjusted according to changes in operating conditions, rather than providing a global solution all at once, thereby maintaining the stability of emission control and the feasibility of operational adjustments under conditions of biomass gas quality fluctuations or load changes.

[0207] Those skilled in the art will understand that the optimization solution method can be implemented using numerical optimization or heuristic search methods suitable for online applications. Its specific form does not constitute a limitation of this application, as long as it can output the required optimization control quantity under the above constraints.

[0208] Next, we will further elaborate on the technical content of the distribution deviation factor in the method of this application.

[0209] It should be noted that deviations under coupled combustion conditions do not always manifest as an increase in the amplitude of a single variable. More commonly, the deviations of multiple mapped variables exhibit synchronous, lagging, or inverse linkages within the same time window, and this linkage changes with the switching of operating conditions. If only the sliding statistics of single-variable deviations are used as the source of uncertainty, the structural information of the deviations will be compressed into a single amplitude range. This means that the safety margin used to generate constraint boundaries can only be widened overall, making it difficult to distinguish which deviations are driven by the same type of source and which are due to the superposition of occasional disturbances. Consequently, the constraint settings may be either too tight, affecting feasibility, or too loose, weakening risk coverage. Therefore, based on the established deviation distribution, an intermediate representation that can express the structural composition of the deviations is introduced, allowing the source contributions of the deviations to enter the subsequent margin generation process in a calculable and updatable form.

[0210] In one example, the step of calculating the difference between the matched measured values ​​and simulation results to obtain the deviation values ​​corresponding to each mapping variable also includes:

[0211] S2.3.1: Within a preset sliding time window, the deviation values ​​corresponding to each mapping variable are arranged into a multidimensional deviation sequence according to the time order, and a deviation dataset is constructed to characterize the time-varying features of the multidimensional deviation sequence.

[0212] Specifically, while the deviation values ​​of mapped variables can reflect the differences between simulation and actual measurements in a single-variable dimension, in the continuous operation of coupled combustion, deviations often appear in the form of multi-variable coordinated changes. For example, deviations in furnace outlet oxygen, staged air distribution, and carbon monoxide or nitrogen oxides may show interrelated change trajectories within the same time period. If each deviation value is treated in isolation, the linkage structure between deviations will be scattered across different statistical calibers, making it difficult to form a unified calculable object.

[0213] In this embodiment, the sliding time window is updated on a rolling basis based on the control cycle. The window length is configured as a combination of several control cycles according to the boiler combustion response and emission monitoring lag, enabling the window to cover the main dynamic process from a single adjustment action to the deviation response. Within each window, the deviation values ​​corresponding to each variable in the mapping variable set are arranged according to the same time index, forming a multidimensional deviation sequence that increases in time. Each moment corresponds to a deviation vector, and each component of the deviation vector corresponds one-to-one with the mapping variable. To ensure comparability between different variables, scaling is performed simultaneously when constructing the deviation dataset: deviation components with large differences in dimensions are standardized or normalized based on the engineering allowable fluctuation range, and the scaling parameters are retained for subsequent reconstruction and write-back. Missing or invalid samples are marked using a consistency mask, enabling the subsequent decomposition process to identify valid dimensions and valid time periods. For short-term missing samples, imputation is allowed within the window. The imputation method can be linear imputation or proximity preservation, and the imputed segment retains a traceable mark in the mask to prevent the shape introduced by imputation from being regarded as the true deviation pattern.

[0214] S2.3.2: Perform sparse factor separation processing on the bias dataset to decompose the bias dataset into multiple latent factor time series and factor loadings corresponding to each mapping variable;

[0215] Specifically, multidimensional deviation sequences are typically formed by the superposition of a small number of dominant change patterns. These patterns exhibit different contribution strengths and directions on different mapping variables. Directly aggregating and statistically analyzing the deviation sequences would lose this contribution structure. However, sparse factor separation can decompose the deviation dataset into several implicit factors and their loadings on each variable without presupposing specific physical source names, allowing the structural information of the deviation to be characterized in an updatable and traceable form. The reason for choosing sparsity constraints is that deviation sources often do not simultaneously and strongly affect all mapping variables. If the loadings are allowed to be uniformly distributed across all variables, the decomposition results are prone to factor semantic aliasing, leading to an unclear contribution structure, which is not conducive to the subsequent formation of distributed deviation factors based on the contribution structure.

[0216] In this embodiment, sparse factor separation takes the bias dataset within a window as input and outputs multiple latent factor time series and factor loadings corresponding to each mapping variable. The latent factor time series describes the intensity changes of each change pattern over time within the window, while the factor loadings describe the strength and direction of each change pattern's effect on different mapping variables. To ensure the decomposition results have stable interpretability, constraints are introduced during the decomposition process: latent factor intensity is subject to a non-negativity constraint, ensuring it can be interpreted as the intensity of bias patterns rather than a superposition of mutually canceling signs; factor loadings are subject to a sparsity constraint, ensuring each latent factor contributes primarily to only a few key mapping variables, preventing excessive factor diffusion; and the latent factor time series are subject to a smoothing constraint and allow abrupt changes, enabling it to track both slowly drifting bias structures and short-term, sudden bias enhancements. The number of latent factors can be determined adaptively, for example, selecting as few factors as possible while ensuring the reconstruction error remains below a preset threshold, or setting an upper limit based on the stability of the number of factors in the historical window, ensuring continuity and computability in online processing.

[0217] S2.3.3: Reconstruct the biased dataset based on the latent factor time series and factor loadings to obtain the reconstructed bias, and calculate the unexplained residual between the biased dataset and the reconstructed bias, wherein the unexplained residual is used to characterize the interpretability of the sparse factor separation to the source of bias;

[0218] Specifically, the latent factors and factor loadings obtained from factor separation need to be reconstructed to verify their interpretability of the bias dataset. Without reconstruction verification, it is difficult to determine whether the bias structure within the current window falls within the coverage of existing change patterns. The reconstructed bias represents the bias portion explained by the latent factors, while the unexplained residual represents the remaining portion that cannot be explained by the latent factors. In an engineering sense, this remaining portion corresponds to the bias components within the window that are not covered by the current factor set. Ignoring this component may lead to an underestimation of uncertainty in the subsequent fluctuation range derived from the bias structure.

[0219] In this embodiment, the bias dataset is reconstructed using the latent factor time series and factor loadings to obtain a reconstructed bias with the same dimension and time index as the original bias dataset. Subsequently, the original bias dataset and the reconstructed bias are subtracted point-by-point under the same index to obtain the unexplained residual sequence. The unexplained residual is used to characterize the interpretability of the sparse factor separation to the source of bias. Its definition employs a computable and comparable metric: one implementation uses the energy percentage of unexplained residuals within a window, representing the residual proportion relative to the original bias dataset; another implementation uses the percentage of unexplained residuals continuously exceeding a threshold within a window, representing the proportion of time the unexplained residuals exceed a preset threshold within the window. These two methods can be used individually or in combination, maintaining consistency in implementation. The threshold can be set based on the residual level of a historical steady-state window or based on the engineering tolerance of each mapping variable, and is updated during the runtime phase.

[0220] S2.3.4: Based on the contribution structure of the reconstruction bias to the corresponding bias values ​​of each mapping variable, and combined with the unexplained residuals, the distribution bias factor is obtained through confidence-weighted quantile statistics;

[0221] Specifically, the formation of the distribution bias factor needs to reflect two types of information simultaneously: one is the bias contribution structure explained by implicit factors, i.e., which implicit factors dominate the biases of different mapped variables and to what extent; the other is the credibility of this explanatory structure, i.e., when the unexplained residuals are large, the stability of the bias structure decreases, requiring more conservative coverage in the statistical summary. Traditional quantile statistics usually directly take quantiles from the original bias sequence, making it difficult to distinguish whether the bias is driven by a stable structural pattern or by unexplained residuals. By combining the reconstructed bias with unexplained residuals for credibility weighting, structural interpretability can be introduced in the statistical summary stage, thus obtaining a distribution bias factor that can be used for subsequent fluctuation range correction.

[0222] In this embodiment, the contribution structure is first determined based on the factor loadings:

[0223] For each mapped variable, calculate the contribution ratio of each latent factor to the reconstruction bias of the variable, and form a contribution weight sequence within the window. The contribution weight can be calculated point by point according to the time step, or it can be summarized according to the average contribution of the window to ensure that the contribution structure is consistent with the bias behavior within the window.

[0224] Furthermore, quantile aggregation is performed using reconstruction bias as the statistical object to obtain a distribution representation of the bias related to the contributions of each latent factor. To introduce confidence, confidence weights are constructed based on unexplained residuals to weight the statistical contributions of different time periods or different sample points within the window: when the unexplained residuals are at a low level, the confidence weights are high, and the quantile statistics are mainly determined by the normal samples of reconstruction bias; when the unexplained residuals increase, the confidence weights decrease, and the upper bound of the bias coverage is amplified during the statistical process so that the distribution bias factor can numerically include the uncertainty expansion reflected by the unexplained residuals. The implementation of confidence weighting can be to assign weights to sample points before performing weighted quantile calculation, or to map the unexplained residuals as an adjustment amount for quantile selection, so that the final quantiles shift towards a more conservative interval.

[0225] In yet another example, the deviation correction sub-model includes factor-wise correction of the fluctuation range based on the distribution deviation factor, specifically including:

[0226] Based on the distribution deviation factor, the fluctuation range of each verification emission within the preset time domain is determined, and the fluctuation range of each sub-factor is weighted according to the contribution weight corresponding to the distribution deviation factor to obtain the comprehensive fluctuation range.

[0227] The overall fluctuation range is reliably corrected based on the unexplained residuals, and the fluctuation range is scaled according to the reliably corrected overall fluctuation range.

[0228] The sub-factor fluctuation range represents the fluctuation range of the verification emission deviation corresponding to a single implicit factor within the preset time domain. The fluctuation range is determined by the component contribution of the implicit factor time series and the factor loading to the reconstruction deviation.

[0229] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins, applied to a production management system, characterized in that... The production management system is equipped with a digital twin model, which includes a mechanism sub-model and a deviation correction sub-model. The mechanism sub-model is used to characterize the material conservation, energy conservation, and pollutant generation mechanisms of the coupled combustion process of biomass gas and coal-fired boiler. The method includes: Acquire sensing data of biomass gas and boiler operation data, and generate state vectors; The state vector is mapped to the simulation results of the digital twin model to obtain the deviation distribution between the state variables and emissions under the corresponding operating conditions. The deviation correction sub-model is then updated based on the deviation distribution to generate emission predictions within a preset time domain. Based on the emission prediction, a constrained optimization problem including emission limits and combustion safety boundaries is constructed and solved to obtain optimized control quantities, wherein the optimized control quantities include at least one of the following: blending ratio, total air volume, and staged air distribution. The state vector is mapped to the simulation results of the digital twin model to obtain the deviation distribution between the state variables and emissions under the corresponding operating conditions, including: Determine the set of mapping variables corresponding to the state vector, wherein the mapping variables include derived state variables for characterizing the coupling relationship between fuel and wind distribution and check emissions for characterizing the emission response; The corresponding measured values ​​are obtained for each mapping variable in the set of mapping variables, and the measured values ​​are matched with the simulation results output by the digital twin model at the same time. The difference between the matched measured values ​​and simulation results is calculated to obtain the deviation value corresponding to each mapping variable. The deviation values ​​are summarized according to preset statistical rules to generate a deviation distribution that characterizes the deviation amplitude and fluctuation range under the target working condition. The statistical rules include at least one of sliding time window statistics and quantile statistics. The deviation correction sub-model is used to generate emission constraint boundaries within a preset time domain based on the deviation distribution, and to generate emission predictions through a prediction algorithm, specifically including: Based on the deviation distribution, the fluctuation range of each verification emission amount within the preset time domain is determined, and the fluctuation range is converted into a safety margin. The safety margin is combined with the preset emission limit to obtain the emission upper limit that varies with the target operating condition, wherein the emission upper limit is less than or equal to the preset emission limit. The emission upper limit is used as the emission constraint boundary, and an emission prediction within the preset time domain is generated by combining the emission constraint boundary with the prediction algorithm. The step of calculating the difference between the matched measured values ​​and simulation results to obtain the deviation values ​​corresponding to each mapping variable includes: Within a preset sliding time window, the deviation values ​​corresponding to each mapping variable are arranged into a multidimensional deviation sequence according to the time order, and a deviation dataset is constructed to characterize the time-varying features of the multidimensional deviation sequence. The bias dataset is subjected to sparse factor separation processing, which decomposes the bias dataset into multiple latent factor time series and factor loadings corresponding to each mapping variable. The biased dataset is reconstructed based on the latent factor time series and factor loadings to obtain the reconstructed bias, and the unexplained residual between the biased dataset and the reconstructed bias is calculated, wherein the unexplained residual is used to characterize the interpretability of the sparse factor separation to the source of bias. Based on the contribution structure of the reconstruction bias to the corresponding bias values ​​of each mapping variable, and combined with the unexplained residuals, the distribution bias factor is obtained through confidence-weighted quantile statistics.

2. The parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins according to claim 1, characterized in that, The sensing data includes the flow rate, temperature, pressure, and lower heating value of biomass gas, as well as the concentrations of the main components corresponding to carbon monoxide, hydrogen, methane, and carbon dioxide. The boiler operation data includes unit load, coal feed rate, primary air volume, secondary air volume, total air volume, furnace outlet oxygen content, furnace negative pressure, main steam flow rate, main steam pressure, main steam temperature, and reheat steam temperature.

3. The parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins according to claim 2, characterized in that, The generated state vector includes: The biomass gas sensing data and the boiler operation data are processed for time synchronization, outlier removal, and dimensional unification. Principal component analysis was used to process the processed sensing data and boiler operation data, and characteristic quantities were selected to characterize combustion boundary conditions, oxygen distribution in the furnace, and heat input characteristics. The feature quantities are combined in chronological order to generate a state vector characterizing the current coupled combustion condition. The state vector includes variables reflecting the energy and composition characteristics of biomass gas and variables reflecting the boiler load and air distribution status.

4. The parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins according to claim 1, characterized in that, The method further includes: The target working condition boundary conditions are determined based on the state vector, and the initial simulation results corresponding to the target working condition boundary conditions are determined by combining the digital twin model. Extract historical operating condition data corresponding to the target operating condition boundary conditions within a preset time window. The historical operating condition data includes at least boiler operation data and actual emission data corresponding to the historical operating conditions. Calculate the deviation between the initial simulation results and the historical operating condition data to obtain the reference deviation; The initial simulation results are corrected by weighted fusion of the reference deviation to obtain the simulation results of the digital twin model. The weighting coefficients are determined by the degree of difference between the boundary conditions of the target operating condition and the boundary conditions of the historical operating condition, and the degree of difference is calculated based on the lower heating value of biomass gas, the concentration of main components and the unit load.

5. The parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins according to claim 1, characterized in that, Based on the emission predictions, a constrained optimization problem is constructed, including emission limits and combustion safety boundaries, and then optimized and solved, including: The weights of the objective function are determined by the analytic hierarchy process to obtain the optimized objective function, which is used to reduce the combustion regulation cost while satisfying the emission constraint boundary. The emission prediction values ​​at each time point within the preset time domain and the corresponding emission constraint boundaries at each time point constitute the emission constraint conditions. Based on the boiler operating data, the combustion safety boundary is determined and the combustion safety boundary is used as a safety constraint condition. The combustion safety boundary includes at least one of the following: the lower limit of oxygen content at the furnace outlet, the allowable range of negative pressure in the furnace, the upper limit of main steam temperature, and the upper limit of reheat steam temperature. The optimized control quantity is determined as a decision variable and the range and rate of change of the decision variable are limited, wherein the decision variable includes at least one of the following: co-firing ratio, total air volume, and staged air distribution. In each control cycle, the optimized control quantity of the previous control cycle is used as the initial value. Combining the emission constraints and the safety constraints, the decision variables are solved in a rolling manner through the optimization objective function to obtain the optimized control quantity of the current control cycle.

6. The parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins according to claim 5, characterized in that, The deviation correction sub-model includes factor-wise correction of the fluctuation range based on the distribution deviation factor, specifically including: Based on the distribution deviation factor, the fluctuation range of each verification emission within the preset time domain is determined, and the fluctuation range of each sub-factor is weighted according to the contribution weight corresponding to the distribution deviation factor to obtain the comprehensive fluctuation range. The overall volatility range is reliably corrected based on the unexplained residuals, and the sub-factor volatility range is scaled based on the reliably corrected overall volatility range.

7. The parameter optimization method for a biomass gas coupled coal-fired boiler based on digital twins according to claim 6, characterized in that, The sub-factor fluctuation range represents the fluctuation range of the verification emission deviation corresponding to a single implicit factor within the preset time domain. The fluctuation range is determined by the component contribution of the implicit factor time series and the factor loading to the reconstruction deviation.

Images