A biomass direct mixing coupled power generation carbon emission prediction method and system

Abstract

This application relates to the field of big data processing, specifically disclosing a carbon emission prediction method and system for biomass direct-hybrid power generation. First, it utilizes aerodynamic flow field to compensate for time lag and achieve mirror alignment of multi-source sensor data. Then, through three-dimensional temperature field reaction zone identification technology, it transforms asynchronous ignition signals into dynamic characteristics and couples them with a twin deep neural network containing built-in ash mass conservation constraints, achieving white-box decoupling of the transient carbon conversion rates of biomass and coal. Finally, based on this, combined with stoichiometric conversion and physical emission limit verification, it completes a forward-looking trend prediction with physical causal terms. This concept fundamentally solves the measurement disconnect caused by asynchronous combustion lag, achieving second-level accurate separation and compliant prediction of dual-source carbon emissions without relying on expensive hardware.

Description

Technical Field

[0001] This application relates to the field of big data processing, and more specifically, to a method and system for predicting carbon emissions from biomass direct hybrid power generation. Background Technology

[0002] Biomass direct hybrid cogeneration of large-scale coal-fired power units has become a key path for the power industry's low-carbon transformation and resource utilization. Because biomass fuels are exempt from carbon emissions under carbon trading policies, while fossil fuel carbon dioxide emissions from coal combustion must be accounted for, the accurate calculation and prediction of scattered carbon emissions under co-firing conditions directly impacts power generation companies' quota revenue and compliance limits, especially given the increasingly stringent carbon emissions trading system. Developing an efficient and accurate carbon emission prediction scheme for biomass direct hybrid cogeneration is not only a prerequisite for supporting refined peak-shaving operation and fuel flexibility management in power plants, but also an urgent need to meet the logical consistency of national carbon monitoring and empower the power industry's deep decarbonization.

[0003] However, existing carbon emission prediction schemes are ill-suited for such complex multi-fuel co-firing scenarios. Current mainstream methods often rely on sensor arrays from continuous emission monitoring systems at the furnace tail end to capture the total physical concentration of carbon dioxide, followed by static proportional mass conversion using simple feed-in feed rates, or end-to-end total emission fitting using black-box models such as long short-term memory networks. These existing technologies suffer from a critical industrial-level problem: in the direct co-firing and stratified combustion of biomass and coal, the bio-based carbon that qualifies for zero-carbon exemption under policy and the fossil-based carbon that needs to be disposed of cannot be dynamically decoupled in real time and accurately predicted. The root causes lie in two aspects. First, the physical limitations of testing methods. In the physical world, CEMS infrared absorption spectrometers can only identify the molecular structure of carbon dioxide, but cannot distinguish in real time whether the carbon atoms originate from fossil coal hundreds of millions of years ago or recent agricultural and forestry waste such as straw. Furthermore, relying on offline isotope sampling and analysis, with its weeks-long cycle, simply cannot meet the second-level real-time requirements of industrial-grade predictions. Second, static mass conversion and purely data-driven models have inherent defects. In configurations with high-proportion coupling (≥10%) such as sidewall layering, biomass has a low ignition temperature (around 200°C) and burns extremely quickly, while pulverized coal has a high ignition temperature (around 400°C), and the time from volatile matter release to coke burnout is not only long but also spatially delayed, resulting in a drastic difference in nonlinear asynchronous thermal conversion between the two. When the boiler undergoes load shaving or sudden feed changes, biomass carbon is instantly generated and reaches the tail end, while fossil carbon continues to burn with a delay in the furnace. Traditional methods are completely blind to this dynamic chemical reaction kinetics, leading to a complete disconnect between the estimated real-time fossil carbon emissions and the physical reality.

[0004] Therefore, an optimized carbon emission prediction scheme for direct hybrid biomass power generation is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for predicting carbon emissions from biomass direct hybrid power generation.

[0006] According to one aspect of this application, a method for predicting carbon emissions from biomass direct-mixed power generation is provided, comprising: S1, based on the flow field aerodynamic delay transmission module, performs spatiotemporal frequency alignment processing on the acquired raw sensor data stream to obtain the aligned combustion state tensor; S2, perform reaction zone identification and asynchronous thermodynamic feature extraction on the aligned combustion state tensor to obtain the dynamic feature matrix; S3, based on the Siamese deep neural network constrained by the embedded fuel ash conservation loss function, performs asynchronous carbon conversion rate decoupling inference on the kinetic feature matrix to obtain the transient carbon conversion rate tuple; S4. Based on the transient carbon conversion rate tuple and the ratio of hysteresis compensation feed and carbon content stored in the aligned combustion state tensor, stoichiometric decoupling calculations are performed on biomass sources and coal sources respectively to obtain decoupled carbon emission data. S5 uses a temporal convolutional sequence network with an attention mechanism to perform trend evolution and forward prediction on decoupled carbon emission data to obtain a carbon emission prediction sequence.

[0007] According to another aspect of this application, a carbon emission prediction system for biomass direct-mixed power generation is provided, comprising: The data spatiotemporal frequency alignment module is used to perform spatiotemporal frequency alignment processing on the acquired raw sensor data stream based on the flow field aerodynamic delay transmission module to obtain the aligned combustion state tensor; The dynamic feature extraction module is used to identify the reaction zone and extract asynchronous thermodynamic features from the aligned combustion state tensor to obtain the dynamic feature matrix. The carbon conversion rate decoupling inference module is used to perform asynchronous carbon conversion rate decoupling inference on the kinetic feature matrix based on the constraints of the embedded fuel ash conservation loss function to obtain the transient carbon conversion rate tuple; The stoichiometric decoupling conversion module is used to perform stoichiometric decoupling conversion on biomass sources and coal sources respectively, based on the transient carbon conversion rate tuple and the ratio of hysteresis compensation feed and carbon element content stored in the aligned combustion state tensor, to obtain decoupled carbon emission data. The trend evolution prediction module is used to perform trend evolution and forward prediction on decoupled carbon emission data through a temporal convolutional sequence network with an attention mechanism to obtain a carbon emission prediction sequence.

[0008] Compared with existing technologies, this application provides a carbon emission prediction method and system for biomass direct-hybrid power generation. Firstly, it utilizes aerodynamics to compensate for time lag and achieves mirror alignment of multi-source sensor data. Secondly, through three-dimensional temperature field reaction zone identification technology, it transforms asynchronous ignition signals into dynamic characteristics and couples them with a twin deep neural network containing ash mass conservation constraints, achieving white-box decoupling of the transient carbon conversion rates of biomass and coal. Finally, based on this, combined with stoichiometric conversion and physical emission limit verification, it completes a forward-looking trend prediction with physical causal terms. This concept fundamentally solves the measurement disconnect caused by asynchronous combustion lag, achieving second-level accurate separation and compliant prediction of dual-source carbon emissions without relying on expensive hardware. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application; Figure 2 This is a data flow diagram of the carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application; Figure 3 This is a flowchart of step S2 of the carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application; Figure 4 This is a flowchart of step S3 of the carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application; Figure 5 This is a block diagram of a carbon emission prediction system for biomass direct hybrid power generation according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] To address the technical challenges in existing multi-fuel direct-hybrid power generation, where the significant temperature difference between biomass and coal ignition and the segmented heat release lag prevent real-time decoupling and accurate prediction of bio-based carbon and fossil fuel carbon emissions, this solution constructs a forward-looking prediction workflow that deeply integrates spatiotemporal dynamics mechanisms with physical laws, thereby effectively predicting carbon emissions from biomass direct-hybrid power generation. In the specific implementation, firstly, an aerodynamic delay transfer model is established to compensate for spatiotemporal delays and align multi-source heterogeneous data such as front-end feed, wind speed, and three-dimensional acoustic temperature field, eliminating data misalignment caused by physical transmission. Subsequently, the aligned three-dimensional temperature cross-section is used for spatial segmentation of the asynchronous ignition field and estimation of mass transfer partial derivatives, accurately extracting the dynamic feature matrix characterizing the heterogeneous heat release rate of the two fuels. Next, this feature is fed into a twin deep neural network in parallel, and a loss function based on the total mass conservation of fuel ash is applied as a strong physical constraint, driving the network to target high volatile matter and high fixed carbon content respectively. The system approximates and decouples high-confidence dual-source transient carbon conversion rates. Based on this, the system dynamically traces and matches historical feed rates with thermal lag and performs stoichiometric conversion based on the carbon element constant-pressure generation ratio, completely separating and forming a real-time decoupled carbon emission baseline sequence. Finally, a time convolutional network with a self-attention mechanism is used to capture the evolution of the fossil carbon flow oscillation slope, and residual clipping verification is performed at the output end with the tailpipe physical discharge and its boundary as rigid monitoring anchor points. This generates a fossil carbon emission prospective prediction map that is physically logically consistent and resistant to variable load interference, completely overcoming the carbon footprint tracking barrier caused by asynchronous combustion of heterogeneous fuels.

[0017] The technical solution of this application proposes a method for predicting carbon emissions from biomass direct hybrid power generation. Figure 1 This is a flowchart of a carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow for predicting carbon emissions from a biomass direct-mixed power generation method according to an embodiment of this application. Figure 1 and Figure 2As shown, the carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application includes the following steps: S1, based on the flow field aerodynamic delay transfer module, the acquired raw sensor data stream is spatiotemporally aligned to obtain an aligned combustion state tensor; S2, the aligned combustion state tensor is subjected to reaction zone identification and asynchronous thermodynamic feature extraction to obtain a dynamic feature matrix; S3, based on a twin deep neural network constrained by an embedded fuel ash conservation loss function, the dynamic feature matrix is ​​subjected to asynchronous carbon conversion rate decoupling inference to obtain a transient carbon conversion rate tuple; S4, based on the transient carbon conversion rate tuple and the hysteresis compensation feed and carbon element content ratio stored in the aligned combustion state tensor, the biomass source and coal source are subjected to stoichiometric decoupling conversion to obtain decoupled carbon emission data; S5: the decoupled carbon emission data is subjected to trend evolution and forward prediction through a temporal convolutional sequence network with an attention mechanism to obtain a carbon emission prediction sequence.

[0018] Specifically, in step S1, based on the flow field aerodynamic delay transmission module, the acquired raw sensor data stream is subjected to spatiotemporal frequency alignment processing to obtain the aligned combustion state tensor. It should be understood that, in the scenario of a 600MW supercritical once-through boiler with a large proportion of coupled co-firing, the coal feeder, biomass feeder, and economizer sensor monitoring the oxygen concentration in the tail flue gas have a physical gas transmission path of tens of meters, and different monitoring devices have different granularity and sampling frequencies. Therefore, in the technical solution of this application, the acquired raw sensor data stream is subjected to spatiotemporal frequency alignment processing based on the flow field aerodynamic delay transmission module to eliminate the causal disconnect of the time axis caused by the spatial transmission lag of the flue gas flow field and the heterogeneity of multi-source sensor sampling. It is worth mentioning that the raw sensor data stream includes the real-time transient inflow rate of the coal feeder and biomass feeder at each level of the furnace, the primary air velocity and secondary air velocity, the three-dimensional temperature cross-section matrix of each level of the burner fed by the acoustic temperature measurement array, and the oxygen concentration in the tail flue gas of the economizer. This ensures that the transient flow rate at the input end, the temperature profile data of each layer in the furnace, and the oxygen concentration effect data at the flue gas outlet are precisely aligned in terms of physical causal logic, providing an aligned combustion state tensor with strong spatiotemporal correlation for subsequent decoupling inference.

[0019] More specifically, in the embodiments of this application, step S1 includes: pulse rejection of real-time transient furnace inlet flow rate, primary air velocity and secondary air velocity based on median filter, and bad zone reconstruction processing of isolated invalid measurement points in the three-dimensional temperature cross-section matrix based on filter matrix to obtain preprocessed sensor data stream; same-frequency resampling mapping of multi-source heterogeneous signals with inconsistent sampling frequencies in the preprocessed sensor data stream within a sliding window to obtain time-frequency data matrix; and time axis remapping of front-end feed data and three-dimensional temperature field data and tail-end exhaust oxygen concentration data in the time-frequency data matrix with total spatial transmission delay as compensation to obtain aligned combustion state tensor.

[0020] Specifically, in one example of this application, the system synchronously acquires raw sensor data streams from the unit control network in real time. These raw sensor data streams include real-time transient feed flow rates of the coal feeders and biomass conveying devices at each level of the furnace, primary and secondary air velocities, three-dimensional temperature cross-section matrices of each burner level fed back by the acoustic temperature measurement array, and flue gas oxygen concentration at the economizer tail. First, the system introduces a dual-dimensional noise filtering layer of temporal filtering and spatial restoration. Based on a median filter, pulse rejection is performed on the real-time transient feed flow rates, primary air velocities, and secondary air velocities to suppress interference from feed jamming or air pressure fluctuations. Then, based on the filtering matrix, isolated invalid measurement points in the three-dimensional temperature cross-section matrix undergo bad zone reconstruction to obtain a pre-processed sensor data stream. Subsequently, the system establishes a global reference timestamp and performs high-order numerical interpolation within a sliding window to resample and map multi-source heterogeneous signals with inconsistent sampling frequencies in the pre-processed sensor data stream to obtain a time-frequency data matrix with a unified discrete grid. Subsequently, the system calculates the total spatial transfer delay under the current combustion condition by integrating along the boiler flue gas dynamic physical trajectory. Using this total spatial transfer delay as a compensation, the system remaps the front-end feed data, three-dimensional temperature field data, and tail-end exhaust oxygen concentration data in the time-frequency data matrix to obtain an aligned combustion state tensor. In calculating the above compensation, the system performs segmented tracking based on the boiler's physical heating surface geometry. The calculation of the total spatial transfer delay in the aerodynamics of the flow field is defined as follows: in, This represents the total spatial delay, expressed in seconds, which is the time span during which the front-end state is transmitted to the location of the tail probe at time t. This represents the total number of flue gas dynamic physical channels that are artificially divided along the airflow axis according to the boiler's physical geometric topology; This represents the equivalent linear axial flue gas flow length within the i-th physical segment, in meters. This represents the average axial velocity of the wet flue gas region, calculated in real-time based on the current transient total air volume, cross-sectional area, and temperature expansion coefficient within the i-th segment at time t, in meters per second. The accumulation operation in the formula decomposes the complex boiler structure into multiple physical segments with independent dynamic characteristics, such as the main combustion segment of the biomass and pulverized coal furnace at an elevation of 13.5m or 17.2m, ascending to the screen-type superheater area, and then extending laterally and descending to the tail economizer flue area. Due to the different flow cross-sectional areas of the different physical segments... While the temperature remains constant, the flue gas temperature decreases gradually due to heat absorption by the heated surfaces within the flue gas. This results in a decrease in the average axial velocity of the locally moist flue gas flowing through each section. It exhibits extremely strong dynamic time-varying characteristics. In this physical scenario, the components... The transient travel time required for an air mass to traverse a cluster of single heat exchange surfaces was precisely quantified, and the results were obtained by summing these values. This method establishes a rigid time-locked anchoring between instantaneous feed abrupt changes or ignition temperature field variations within the furnace and residual oxygen changes detected by the flue gas probe after a specific delay. By shifting the total delay compensation amount back to the historical timeline and tracing this aerodynamic space, it successfully restores the strict cause-and-effect concurrency pattern of the front-end and rear-end variables within the aligned combustion state tensor on the logical slice, thus eliminating the potential correlation disaster that could arise from directly using raw meat data for black-box training at the physics level.

[0021] Specifically, in step S2, the reaction zone is identified and asynchronous thermodynamic features are extracted from the aligned combustion state tensor to obtain the kinetic feature matrix. It should be understood that in the direct coupling and co-firing scenario of large-scale coal-fired units with a large proportion of sidewall-layered biomass burners, the biomass fuel fed into the furnace has extremely high volatile matter and an extremely low ignition temperature, with its critical ignition reaction point only between 200 and 260 degrees Celsius and overall combustion being extremely rapid. In contrast, the normal pulverized coal fed into the furnace by the front and rear wall burners has an ignition temperature as high as 350 to 420 degrees Celsius and requires an extremely long high-temperature coke burnout stage. These two fuels with vastly different physical properties generate extremely intense and disparate asynchronicity in heat release and spatiotemporal heterogeneity within the same furnace mixing and heat exchange space. If the entire boiler combustion area is considered as a unified, single physical black box field, it will inevitably lead to severe signal aliasing in the oxygen consumption patterns of the misaligned distribution of different intrinsic fuels. Therefore, in the technical solution of this application, the reaction zone is further identified and asynchronous thermodynamic features are extracted from the aligned combustion state tensor to obtain a kinetic feature matrix. This matrix is ​​used to decouple the asynchronous ignition field in three-dimensional space and separately extract the dynamic gradient attributes characterizing the actual oxidation reaction rates of the two heterogeneous fuels. In this way, the calculation interference caused by the spatial overlap of dual-source fuel co-combustion can be completely eliminated, and the complex and disordered multi-source physical combustion behavior can be reduced in dimension and mapped into a kinetic feature expression with a high signal-to-noise ratio. This provides pure thermodynamic boundary constraints and underlying structured feature support for subsequent deep digital model execution of independent and high-confidence real-time carbon conversion rate stripping inference.

[0022] Figure 3 This is a flowchart of step S2 of the carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application. Figure 2 As shown, according to an embodiment of this application, step S2 includes: S21, dividing the three-dimensional temperature cross-section data in the aligned combustion state tensor into a spatial reaction zone of the asynchronous ignition temperature field to obtain a spatially segmented combustion tensor; S22, estimating the combustion mass transfer partial derivatives and heterogeneous time constants of the spatially segmented combustion tensor to obtain a heterogeneous time constant feature set; S23, performing dimensionality reduction and compression encoding on the heterogeneous time constant feature set to obtain a kinetic feature matrix.

[0023] Accordingly, in step S21, the spatial reaction zone of the asynchronous ignition temperature field in the three-dimensional temperature section data of the aligned combustion state tensor is segmented to obtain the spatially segmented combustion tensor. It should be understood that due to the essential differences in the inherent physicochemical properties of biomass and pulverized coal, specifically, biomass particles injected into the furnace from the side wall ignite very early and rapidly, with their initial volatilization and heat release zone highly concentrated in a specific medium-low temperature gradient range of 200 to 350 degrees Celsius, while the main pulverized coal particle group injected into the furnace from the front and rear walls needs to cross an extremely high reaction activation energy barrier, and its core high-temperature coke main combustion zone is absolutely in a localized violent swirling field above 400 degrees Celsius. If these two objectively adjacent regions with vastly different heat release temperature boundaries cannot be precisely and exclusively segmented and defined in the extremely complex physical three-dimensional space, it will inevitably lead to irreversible dual-source feature aliasing and heterogeneous kinetic data contamination when calculating the response law of oxygen consumption rate fluctuation to feed changes for arbitrary spatial cells. Therefore, in the technical solution of this application, a spatially segmented combustion tensor is obtained by asynchronously dividing the spatial reaction region of the ignition temperature field in the three-dimensional temperature cross-sectional data of the aligned combustion state tensor. This allows for the explicit stripping, shielding, and assignment of exclusive physical affiliation labels to independent three-dimensional grid coordinates belonging to different combustion release characteristics from the macroscopic multidimensional thermodynamic temperature matrix, based on clear thermophysical threshold boundaries. This ensures that downstream algorithms have a completely pure and non-interfering spatial physical domain as their computational foundation when extracting their respective transient reaction time delays and calculating partial derivatives. This cuts off the coupling noise caused by chaotic multi-fuel co-firing at the root of the data structure, resulting in a tensor with a spatially binarized weighting attribute with an extremely high signal-to-noise ratio.

[0024] More specifically, in the first embodiment of this application, the system extracts the three-dimensional temperature cross-sectional data of the holographic structure from the aligned combustion state tensor output by the pre-spatiotemporal compensation algorithm. For the two fuels with completely different thermal volatilization decomposition and coke oxidation initiation boundaries, a hard step indicator function is introduced to execute a binary determination strategy per spatial cell. The system sets up dual, non-intersecting physical temperature filtering pipelines to lock and assign a value of one to spatial cells within the predetermined biomass reaction range in the three-dimensional coordinate system, while forcibly setting the remaining regions to zero. Parallel and independently, regions with temperatures exceeding the stable coke combustion framework temperature are locked and assigned a value of one, while others are shielded. The determination results based on hard gradients in the Cartesian coordinate system jointly construct a pair of orthogonal trimask arrays with independent spatial mapping labels. These mask arrays perform point-to-point retention or truncation and erasure of the full gas flow, feed, and temperature field variables in the source tensor, ultimately encapsulating and combining them to output a spatially segmented combustion tensor carrying precise spatial segmentation authority. In the process of implementing the hard segmentation of the spatial reaction region based on cell-by-cell binarization, the system internally constructs a rigorous set of step indicator mask functions to establish the dual-source reaction mask: Where x, y, and z represent discrete three-dimensional Cartesian coordinate scalars of fixed width in the physical heat exchange area of ​​the furnace that are digitized by the system; t represents the parameter at the same global synchronization moment when the system performs local instantaneous detection, judgment, and segmentation calculation. This represents the coordinate system based on three-dimensional temperature cross-section data. The local instantaneous absolute physical temperature value extracted at the t mark is provided in degrees Celsius. This indicates a hard step indicator operator function with strict exclusive logic properties. It returns a real 1 when the physical comparison condition inside its parentheses is true, and a real 0 otherwise. This represents a Boolean mask for the biomass reaction zone, generated by the system's computational deduction, used to characterize whether the grid site belongs to the influence domain of the preceding biomass forced-release combustion at that moment. This represents a Boolean mask for the coal reaction zone generated by the system to determine whether it fully belongs to the subsequent high-potential coal-to-electricity conversion reaction domain. The defined boundary condition range precisely captures the core temperature zone occupied by the highly volatile components of biomass composed of a large amount of agricultural and forestry waste lignin and cellulose, which, once detached from the nozzle and heated in the furnace to absorb convective heat energy, undergo explosive decomposition first. Any physical grid entering this parameter range, after... After the function operation, the system is forced to output state 1, forming a digital shell that thoroughly delineates the influence boundary of non-fossil gas masses. Correspondingly, for pulverized coal fuel flows that are structurally dense, highly metamorphic, and lack rapid combustion activity, the system sets up a one-way physical checkpoint. Its significance lies in ignoring the ineffective preheating stage of slow initial heat absorption by pulverized coal, directly and precisely locking onto and cutting out the high-temperature radiative core array where violent solid-phase coke core oxidation occurs, and then outputting a binary numerical response mask via an indicator function. The two conditional expressions above each output... and By directly performing binary isolation and cutting on the three-dimensional furnace, an absolute zero-tolerance transition and an ambiguous hard step shielding mechanism were successfully created. This constitutes the most basic and direct framework for realizing the separation, decoupling, and qualitative judgment of dual-source multi-domain physical layer features in the early stages of the carbon footprint monitoring system construction.

[0025] In particular, in the carbon emission prediction method for biomass direct-hybrid power generation, the identification and extraction of the reaction zone with asynchronous thermodynamic characteristics is a key upstream link in the entire carbon decoupling prediction chain. The core task of this link is to spatially segment the data of the biomass combustion zone and the pulverized coal combustion zone from the aligned combustion state tensor, so that subsequent mass transfer partial derivative calculations and twin network inferences can independently perform carbon conversion rate decoupling for the two fuels.

[0026] The first embodiment described above uses a hard step indicator function to perform cell-by-cell binarization of the three-dimensional temperature cross-section field. Specifically, cells with temperatures between 200°C and 350°C are marked as the strong heat release zone of biomass, while cells with temperatures exceeding 400°C are marked as the main combustion zone of pulverized coal. While mathematically simple, this scheme reveals a structural defect under engineering field conditions that is mismatched with the depth of the sidewall stratified combustion configuration.

[0027] It is worth mentioning that this 600MW supercritical once-through boiler adopts a layered arrangement of biomass burners on the side walls. The upper layer of biomass burners is located at an elevation of 17.2m between the middle and upper layers of pulverized coal burners, while the lower layer of biomass burners is located at an elevation of 13.5m between the middle and lower layers of pulverized coal burners. A total of 24 pulverized coal burners are arranged in three layers at similar elevations on the front and rear walls. The direct physical consequence of this spatial layout is that the biomass combustion zone on the side walls and the pulverized coal combustion zone on the front and rear walls are spatially adjacent at the same elevation level. The high-temperature radiation fields generated by the two fuels undergo significant thermal radiation cross-coupling in the central region of the furnace, and the hard threshold binarization segmentation of the first embodiment completely obscures this special physical relationship.

[0028] Specifically, the first manifestation of the defect is the overall loss of transition zone data due to the temperature blind zone. The original hard threshold mask creates a 50°C-wide judgment vacuum zone between the upper limit of biomass judgment (350°C) and the lower limit of pulverized coal judgment (400°C). The spatial cells within this temperature range are neither classified into the biomass reaction zone nor the pulverized coal reaction zone, and are directly assigned a value of zero and discarded in the mask calculation. However, in the sidewall layered configuration, this temperature transition zone corresponds precisely to the central space of the furnace where the thermal radiation of the two fuels most actively converges and overlaps. Taking the 17.2m elevation level as an example, the biomass particles injected by the sidewall biomass burner ignite rapidly and release volatiles in the 200°C~350°C range, while the pulverized coal injected by the front and rear wall pulverized coal burners forms a stable high-temperature coke combustion core zone above 400°C. The radiant energy of the two superimposed in the intersection area of ​​the furnace center forms a mixed temperature field of 350°C~400°C. This region carries crucial transitional information regarding heat exchange and mass transport between the two fuels. The first embodiment discards this information entirely, resulting in a severe loss of upstream information. This directly leads to a deterioration in the accuracy of subsequent mass transfer partial derivative calculations and a decrease in the confidence level of carbon conversion rate decoupling inference.

[0029] The second manifestation of the defect is the unattributable nature of radiation cross-contributions. Even if the temperature threshold of the transition zone is artificially widened to bridge the blind spot of 350°C to 400°C, the hard binary mask can only classify each spatial cell as either entirely biomass or entirely pulverized coal. However, in the reality of furnace physics, the measured temperature of any spatial cell in the transition zone is the combined result of the superposition of dual-source thermal radiation from adjacent biomass and pulverized coal combustion zones. This measured temperature contains the radiation contribution fractions of each fuel, the proportion of which depends on the spatial distance between the cell and the core areas of the two reaction zones, as well as their respective radiation intensities. The 0 or 1 logic of hard binary determination is simply incapable of expressing this continuous fractional physical attribution relationship, causing the carbon source allocation in the transition zone to deviate from physical reality.

[0030] The third manifestation of the defect is the gradient noise contamination of downstream differential operations caused by the step discontinuity. The hard step indicator function produces an abrupt jump in the mask value from 0 to 1 at the temperature threshold point. This discontinuity is amplified into a spurious spatial gradient pulse in the subsequent mass transfer partial derivative differential calculation step. This is because the differential operation is essentially an approximation of the change between adjacent cells using finite difference quotients, and the artificial step of the mask is unrelated to the actual combustion physics gradient. The resulting pseudo-gradient noise directly degrades the signal-to-noise ratio of the kinetic feature matrix, and then propagates along the data stream to the decoupling inference stage of the twin network, ultimately affecting the accuracy of carbon emission prediction.

[0031] In summary, the hard threshold binarization segmentation of the first embodiment, in the scenario of large-scale biomass co-firing with sidewall layered arrangement, fails to consider the inherent thermal radiation cross-coupling attribution relationship between the reaction zones of the two fuels, resulting in the triple superposition of data loss in the transition zone, lack of source attribution, and downstream noise amplification, which becomes a key bottleneck restricting the improvement of carbon emission prediction accuracy.

[0032] To address the aforementioned shortcomings, an improved mechanism is proposed. This mechanism employs a weighted soft computation framework with gradient-triggered smooth decay and an embedded Stefan-Boltzmann spatial four-way radiation competition mechanism. It flexibly and continuously prunes the transition dead zone based on physical causal correlation and a fractional allocation mechanism. This effectively severs the implicit information loss caused by temperature blind spots and the chain reaction of step noise caused by physical truncation from the very upstream of the data link. It transforms the subjective blind box, based solely on empirical temperature judgments, into a high signal-to-noise ratio smooth soft tensor analytical foundation grounded in the quantification of causal tracing of optical transmission and heat exchange. This provides a multi-dimensional data environment with both high fidelity resolution and strong robustness for subsequent twin model separation of bidirectional dynamic characteristics.

[0033] More specifically, in the second embodiment of this application, the spatial reaction zone of the asynchronous ignition temperature field in the three-dimensional temperature cross-section data of the aligned combustion state tensor is segmented to obtain a spatially segmented combustion tensor, including: performing gradient adaptive continuous soft membership estimation on the aligned combustion state tensor to obtain an adaptive soft membership tensor; performing cross-coupling attribution correction on the adaptive soft membership tensor to obtain a radiation-compensated membership tensor; and performing weighted soft segmentation based on compensation membership on the radiation-compensated membership tensor and the aligned combustion state tensor to obtain a spatially segmented combustion tensor.

[0034] Specifically, gradient adaptive continuous soft membership estimation is performed on the aligned combustion state tensor to obtain an adaptive soft membership tensor. It should be understood that the original hard threshold mask, using a globally fixed temperature boundary for spatial segmentation during calculation, completely ignores the local differences in the spatial gradient distribution of the temperature field inside the furnace. Under the specific physical reality of the sidewall layered configuration, the temperature gradients within the core regions of both the biomass combustion zone and the pulverized coal combustion zone are extremely steep, a characteristic directly caused by the heat accumulation effect of the swirl burner; while the temperature gradient in the central transition mixing zone between the two reaction zones is relatively gentle, a characteristic caused by the diffusion effect of heat radiation and heat exchange. If the same rigid segmentation bandwidth is applied to these two distinctly different physical spatial regions, either boundary ambiguity will occur in the core region, leading to fuel characteristic contamination, or excessive truncation will occur in the transition zone, creating an artificially isolated information zone, making it impossible to simultaneously achieve targeted segmentation accuracy and the continuity of the mixing zone transition. Therefore, in the technical solution of this application, gradient adaptive continuous soft membership estimation is further performed on the aligned combustion state tensor to obtain an adaptive soft membership tensor. This is used to feed back the local temperature change rate to the truncation bandwidth boundary, realizing a spatial adaptive segmentation strategy of sharp segmentation of the core region and flexible transition of the mixed transition region. Furthermore, a composite bidirectional continuous activation function framework is used to coherently assign probabilistic weights to the array mesh with blind zones. In this way, the problem of overall loss of the underlying data of the transition region caused by the temperature blind zone of 350 to 400 degrees Celsius can be eliminated, while maintaining the segmentation accuracy of the source features of the core combustion region without any damage. Moreover, thanks to the global continuous differentiability of the smoothing function, the fatal problem of pseudo-gradient differential noise propagating to the downstream mass transfer partial derivative difference operation network is avoided from the root of physical calculation.

[0035] In a specific example of this application, the system sequentially executes a series of core computational steps, from spatial gradient identification and capture to smooth tensor encapsulation and reconstruction. First, the system analytically extracts the three-dimensional holographic temperature cross-sectional field from the accessed aligned combustion state tensor, and calculates the three-dimensional spatial gradient magnitude of the temperature field at each spatial cell within the global coordinate grid, thereby directly characterizing the drastic degree of temperature field evolution at different spatial locations. The physical derivative analytical formula used by the system to calculate the spatial gradient magnitude of the temperature field is: in, To calibrate the spatial coordinates of the furnace The local temperature scalar value at time t, in units of , Let Euclidean norm be the three-dimensional spatial gradient of the temperature field, in units of 1000 kJ / m². The physical meaning of this value is the intensity of the spatial rate of temperature change at the location of the cell. The three partial derivative terms within the equation precisely and accurately capture the slope of the heat flow temperature fluctuations along the three orthogonal geometric axes of width, depth, and height. The norm value obtained by summing the squares and taking the square root of these terms equivalently maps the overall intensity of the rate of thermal change at the pose of the cell point. This value exhibits a huge peak at the edge of the flame core subjected to radial swirling flow from the burner, while showing an extremely low characteristic in the central large-space mixing region where various radiant heats are gradually depleted and dispersed, providing an extremely accurate and objective kinetic physical criterion for the generation of adaptive interfaces.

[0036] Subsequently, the system constructs an adaptive smoothing bandwidth parameter for each spatial cell based on the extracted gradient magnitude. Under the control of this parameter, the system automatically narrows the bandwidth in the core combustion zone with a steep temperature gradient to strictly maintain the segmentation sharpness and avoids erroneously blurring and diffusing the original distribution data of the core high-temperature zone belonging to a single dominant fuel to the belonging space of another weak thermal potential fuel. At the same time, in the intermediate transition zone with a gentle temperature gradient, the system automatically widens the bandwidth to both sides to allow for a smooth cross-transition of continuous membership degrees for the coexistence of dual fuel thermal fields, thereby eliminating the temperature determination dead zone and the resulting overall data loss problem at the source of the data flow. The system constructs the adaptive smoothing bandwidth boundary formula for each cell by combining Euclidean norm values ​​as follows: in, For spatial cells The adaptive temperature smoothing bandwidth at time t, in units of , This is the initial constant for the global reference bandwidth, determined by engineering experience or offline calibration, with a typical value range of [value missing]. to , The bandwidth attenuation sensitivity coefficient controls the response of the gradient magnitude to bandwidth contraction and is a dimensionless positive real number. Combined with the internal spatial distribution of the irregular furnace configuration, this fractional mechanism achieves perfect suppression and relief. When the grid computation node enters the near-field of the swirling core combustion, the gradient norm carried by the denominator undergoes a huge expansion and rise, forcing the fraction to be rapidly swallowed and suppressed by the denominator, resulting in a highly convergent and extremely low bandwidth surface, strictly preventing signal overflow contamination. Conversely, when the computation node falls back into the transitional flow region of the central cold zone, the norm term weakens, and the denominator approaches a constant of one, preserving the broad and relaxed initial constant bandwidth scale, like a wide net recaptures the crudely missed fusion attenuation data.

[0037] It is understandable that the spatial distribution of the temperature field at different elevation levels in the furnace within the layered sidewall configuration is extremely uneven. The temperature gradient near the biomass burner nozzle on the sidewall can differ from the temperature gradient in the swirl core area of ​​the pulverized coal burner on the front and rear walls by several times, while the gradient in the central transition zone between the two is significantly lower than that in any core zone. By feeding the gradient information back to the bandwidth parameter, a spatial adaptive segmentation strategy is achieved, characterized by sharp segmentation of the core zone and flexible transition of the transition zone.

[0038] Subsequently, the system utilizes the generated adaptive bandwidth scale and, through a decision mechanism that replaces the original hard step indicator function with a logarithmic sigmoid continuous function, constructs continuous soft membership degrees for both the biomass reaction zone and the pulverized coal reaction zone. Specifically, the biomass membership degree employs a bandpass double sigmoid composite structure to characterize 200... Up to 350 The smooth transition characteristics within the temperature range, and the membership of pulverized coal were characterized using a high-pass single Sigmoid structure at 400°C. The smooth activation extension characteristics near the high-energy potential threshold trigger point. The calculation formula for the combined distribution system of smooth and coherent soft membership curves of biomass and coal is as follows: in, For spatial cells The continuous soft membership degree of the biomass reaction zone, i.e., the probability weight of membership matching attribution, takes the value of a continuous real number between 0 and 1. For spatial cells Continuous soft membership of the pulverized coal reaction zone For the standard Sigmoid function, The lower limit temperature threshold for the biomass reaction zone is set to 200. , The upper temperature threshold for the biomass reaction zone is set at 350°C. , The lower limit temperature threshold for the pulverized coal reaction zone is set at 400. , This is the adaptive smoothing bandwidth calculated above. Within the transition temperature range, and Each outputs a consecutive non-zero value (e.g., approximately 0.3 to 0.5 each), thereby discarding the 350 values ​​that were completely discarded in the first embodiment. Up to 400 The transition zone data is recycled into the segmentation results in the form of fractional weights. From the perspective of integrating mathematical structure with actual physical processes, this composite formula completely dismantles the crude black-and-white demarcation of the past. In the biomass formula, the two opposing functions interfere with each other and multiply, forming a naturally smooth bandpass with a rising edge connected to a smooth base and a falling edge, gently but precisely enveloping the extremely short-lived main temperature range of volatile combustion of agricultural and forestry waste; the arrangement of the unidirectional high-pass function in the coal formula conforms to the divergent nature of the long tail of coke's continuous high heat. Through the synergistic effect of gradient-driven adaptive bandwidth mechanism and continuous Sigmoid function, the 350°C bandpass is completely eliminated. Up to 400 While addressing the data loss issue in the transition zone caused by the temperature blind zone, the segmentation accuracy of the core combustion zone remains intact. Furthermore, due to the continuous differentiability of the Sigmoid function, the problem of pseudo-gradient noise propagating downstream mass transfer partial derivative difference operations due to hard step discontinuities is fundamentally avoided. Finally, the continuous soft membership of each spatial cell, along with the temperature field data at its corresponding location, is encapsulated into an adaptive soft membership tensor.

[0039] Specifically, the adaptive soft membership tensor is cross-coupled and attribution-corrected to obtain the radiation-compensated membership tensor. It should be understood that while the pre-set temperature gradient-based soft membership evaluation mechanism addresses data loss and step noise in the transition zone through continuous soft membership, it only uses the temperature scalar of a single spatial cell for local evaluation, without considering the physical contribution of deep thermal radiation energy transfer between adjacent reaction zones in the complex three-dimensional furnace system to the temperature formation of that specific cell. Especially in large-scale coupled configurations such as sidewall layering, the biomass burner on the sidewall and the pulverized coal burners on the front and rear walls are extremely adjacent at the same physical elevation, directly causing the measured temperature of the transition zone grid cell to be not simply the heat release from combustion of a single pure fuel, but rather a composite result of the superposition of radiative energy projected onto the cell from completely different spatial orientations by these two heterogeneous fuels. Isolated temperature values ​​alone are insufficient to accurately distinguish the exact proportion of contribution from biomass radiation and the specific proportion from pulverized coal radiation within the mixed temperature system. Therefore, this application's technical solution further refines the adaptive soft membership tensor through cross-coupling attribution correction to obtain a radiation-compensated membership tensor. This allows for the in-depth introduction of a spatial causal attribution mechanism based on the physical laws of radiation transfer to comprehensively correct the initial soft membership weights. By quantifying the energy projection competition state of each radiation source, and without adding any additional hardware sensors, the existing dynamic temperature field data system can be used to achieve precise and quantitative attribution of the physical contributions of dual-source thermal radiation in the aliasing transition zone. This elevates the system's judgment of fuel space segmentation from an empirical reliance on shallow temperature values ​​to an objective causal attribution based on the physical laws of deep thermal radiation transfer, fundamentally eliminating the blindness to the cross-coupling effect of dual-source thermal radiation under a layered configuration with large-scale co-firing sidewalls.

[0040] In a specific example of this application, the system starts from the fundamental law of radiative transfer physics—that the absolute emission intensity of the self-generated thermal radiation from the surface of any high-temperature object is directly proportional to the fourth power of its absolute thermodynamic temperature—and performs a correction for the adaptive soft membership tensor. In engineering physics, this law implies that although the temperature of the main combustion core region is often around 1200°C... The radiation emission intensity in the above-mentioned pulverized coal combustion zone far exceeds that at a temperature of 200°C. Up to 350 The biomass heat release zone, however, often occupies a much larger volume coverage area in the three-dimensional transition space, creating a power imbalance between the concentrated high-temperature coal jet and the extensive volume coverage of the biomass scattering surface, resulting in a weighted game of radiative influence. Based on this objective physical principle, the system first calculates the cumulative radiative influence kernel function for each spatial cell requiring physical source attribution, independently and separately, for all effective source cells from the biomass and pulverized coal reaction zones. This radiative kernel function strictly follows the intensity weight distribution criterion of the Stefan-Boltzmann fourth-power radiation law and introduces an exponential spatial attenuation factor to describe the nonlinear decrease in radiation intensity as it travels through particulate matter and flue gas within the furnace. Within this correction calculation step, the system defines the formula for the cumulative radiative influence kernel function used to calculate the long-distance ejection of various specified fuel spatial zones to specific target cells as follows: in, For time t, from the reaction zone All spatial cells to target cell The cumulative radiative impact generated, in units of This reflects the equivalent order of magnitude of the radiation energy density transmitted to that location. The physical coordinate geometric set of all discrete spatial origin cells within the three-dimensional effective combustion space domain that covers the entire furnace interior is defined as the entire effective combustion space domain. For source cell Belongs to the reaction zone Continuous soft membership weights The thermal radiation emission intensity of the source cell is represented by the fourth power of the source cell temperature, following the Stefan-Boltzmann radiation law. This explicitly calibrates and exposes the fundamental physical emission capability of the source mass, which exhibits a strong and compelling external thermal radiation. The photon absorption attenuation constant coefficient is the medium's ability to strongly intercept and reduce photonuclear thermal ripples when a mixture of densely suspended coal particles, disordered biomass fly ash, and highly polluted thick smoke plumes crosses a heavily polluted flue gas field. Its unit is [unit missing]. This reflects the absorption and shielding effect of high concentrations of pulverized coal particles, biomass fly ash, and flue gas on radiation within the furnace. The distance between the target pose point of the receiver cell and the location point of each microscopic source in space is the direct Euclidean distance vector, expressed in meters (m). For a cell in a transition zone at the center of the furnace at an elevation of 17.2m, both the lower-temperature but larger-volume radiation source group from the biomass burner along the sidewalls and the extremely high-temperature but relatively concentrated radiation source group from the pulverized coal burners along the front and rear walls simultaneously project radiant energy onto the cell. The competitive relationship accurately describes the physical causal structure of the temperature formation in this lattice.

[0041] Subsequently, the system utilizes the calculated dual-source radiation influence kernel function, combined with the initial continuous soft membership degrees issued in the preceding stage, to perform cross-attribution fusion and iterative correction. During this process, the correction algorithm innovatively incorporates a competitive normalized fractional structure, which combines the product of the node's original initial membership degree and the corresponding calculated cross-boundary radiation influence value as the allocation numerator vector. Simultaneously, it merges the sum of the joint products from two sources—natural biomass and fossil coal—as the denominator array base, thereby rigorously performing targeted normalized competitive proportion allocation calculations. Finally, the system uniformly encapsulates and synthesizes the final compensated membership degree data of each reaction zone obtained through rigorous cross-energy game judgment into a radiation compensation membership degree tensor. To inject physical competition constraints into the original fuzzy probability after obtaining this perspective-oriented influence index, the competitive normalized fusion balanced algorithm is as follows: in, For spatial cells after attribution correction for radiative cross-coupling The final compensated membership degree for reaction region k takes the value of a continuous real number between 0 and 1. This indicates that the underlying algorithm, designed for protection against competitive pooling, is highly susceptible to collapse in environments such as fireless suspension, extreme cold, deep currents, or sudden shutdowns. This could trigger a complete system crash, causing the denominator to plummet and the division function to fail. The system's defensive capabilities include minimal positive real numbers, singular normal values ​​(typically 10). -8 This is used to prevent numerical singularities where the denominator is zero. If a probe located at a neutral point extremely far from the combustion wall, for example at a height of 17.2m, provides a feedback value that is fixed at 375... When faced with such temperatures, the spatial cell of a temperature point will be mercilessly trapped in an ambiguous and unclear origin zone if its location is significantly close to the high-potential energy center of continuously burning coal powder at full power on the front and back walls, and subjected to extremely intense thousand-degree heat and nuclear explosion, resulting in the flux influenced by coal radiation. It demonstrated an absolute, overwhelming, and disproportionately large-scale defeat of the lignin combustion stream originating from the extremely loose, low-heat-emitting volume on the opposite side. At that time, the results obtained by calculating the stretching and folding through this set of physical causal equations were... The group's weights will exhibit an astonishing surge in absolute power and an undeniable crushing dominance. This quantitative comparison phenomenon, directly induced by physical competition, acts like a detector, precisely clarifying to the deep calculation framework the truth that although the measured temperature of this segment is moderate, its causal background is entirely dominated and manipulated by the externally mounted high-potential coal nucleus long-range radiation wave. Conversely, in contrast, if the suspended fixed point of this same temperature reading segment were extensively expanded and diffused, resulting in a large, flat area covering the sidewall combustion zone and surrounding the biological area, and if the calculated radiative nuclear source kinetic energy... If a stock rises to its peak and seizes control of the main trend, then the aforementioned equilibrium calculation system will definitely allow it to... The dominant control takes precedence. Undoubtedly, this mathematical compensation and recalculation system, built upon extremely rigorous physical transmission laws and fractional dynamic occupancy weight reconstruction, breaks through the inherent constraints and short-sighted judgment problems of blindly applying surface temperature readings. It establishes a lossless spatial rights confirmation baseline that restores the objective and true laws of energy waveform transmission through walls, comprehensively preventing and resolving the problem of data contamination and defocusing caused by the mutual interference of radiation cross-layers. Correspondingly, by introducing the Stefan-Boltzmann fourth-power radiation law and the spatial exponential decay kernel function into the reaction zone segmentation judgment, without adding any hardware sensors, it can achieve quantitative attribution and decomposition of the dual-source thermal radiation contribution in the transition zone using only existing temperature field data. This elevates the segmentation result from empirical judgment based on temperature values ​​to causal attribution based on the physics of radiation transmission, fundamentally eliminating the blindness to radiation cross-coupling effects in the first embodiment.

[0042] Specifically, a weighted soft segmentation based on the compensated membership degree is performed on the radiation-compensated membership tensor and the aligned combustion state tensor to obtain a spatially segmented combustion tensor. It should be understood that although the system has generated gradient-adaptive soft membership degrees and radiation-attribution-corrected compensated membership degrees in the first two stages, these membership degrees are merely abstract and independent scalar weight parameters. They have not yet been substantially combined with the underlying data stream carrying the real dynamic physical mapping, and have not been applied to the full amount of multi-channel original combustion state data contained in the aligned combustion state tensor. This results in the multidimensional original tensor remaining unweighted by physical boundary assignment, unable to construct a standard heterogeneous input format that can be directly received and analyzed by the subsequent mass transfer partial derivative calculation steps. Therefore, in the technical solution of this application, a weighted soft segmentation based on the compensated membership degree is further performed on the radiation-compensated membership tensor and the aligned combustion state tensor to obtain a spatially segmented combustion tensor, thereby transforming the pure probability scalar into a mathematical multiplicative channel filter operator capable of thoroughly decomposing the characteristics of multi-source complex mixed flow fields. In this way, the physical attribution conclusions of radiation cross-coupling can be successfully imprinted into the underlying data volume covering all measurement dimensions such as temperature, wind speed, and feed. Compared with the first embodiment, which can only produce an extremely coarse segmentation of all or nothing based on hard binary masks, the improved weighted soft segmentation fully preserves the complete information matrix of the transition region and realizes the complete assignment and delimitation of the fractional true components based on the differentiation theory of radiation dynamics.

[0043] In a specific example of this application, the system performs a data multiplication, combination, recombination, and distribution operation on the global furnace spatial elevation plane. The system first uses the radiation cross-compensation membership coefficients specific to biomass and coal, contained in the radiation compensation membership tensor, as continuous mapping weight scalars for each spatial geometric cell. Subsequently, the system substitutes these continuous weight coefficients point-to-point into and extracts the full combustion state data carried by each corresponding spatial cell in the aligned combustion state tensor. This full combustion state data comprehensively encompasses all physical detection vector information across all channel dimensions, including the absolute temperature field, dynamic bias airflow velocity, and mass feed conversion, accompanying a specific spatial point at any given time. After completing the high-throughput weighted extraction and separation point-to-point and channel-to-channel, the system then performs a layered splicing and recombination of the corresponding biomass attribute vector components and corresponding coal powder attribute vector components derived from all spatial grid cells according to the strict original spatial coordinate dimensions, ultimately sealing them into a complete spatially segmented combustion tensor with pure and independent dual-channel reaction zone physical attribution identifiers.

[0044] In the data extraction and spatial tensor reconstruction stage, the core matrix product function model for performing weighted soft-axis splitting on the aligned multi-channel physical vector dimension is as follows: in, This indicates a specific classification code representing the individual fuel stripping characteristics, with its value range representing the biomass reaction zone. System and representative physical coal-fired reaction zone system, This is the weighted segmented combustion state data vector belonging to reaction zone k after radiation attribution weighting. This is the radiative cross-coupling compensation membership scalar derived from the fixed-frequency output of the pre-processed single-point radiative cross-coupling correction step. To align the combustion state tensor in the space lattice The original multi-channel combustion state data vector was then generated. Subsequently, the biomass and pulverized coal components of all spatial cells were stacked and recombined according to the spatial coordinate dimensions, and encapsulated into a complete spatially segmented combustion tensor with dual-channel reaction zone attribution labels.

[0045] Compared to the hard binary mask of the first embodiment, which can only produce a coarse segmentation with all or nothing results, the improved weighted soft segmentation retains complete information about the transition region and achieves fractional membership based on radiation physics. Furthermore, due to the continuous differentiability of the entire link, no spurious spatial step gradient noise is introduced when transferring to the subsequent mass transfer partial derivative difference calculation step, ensuring the signal-to-noise ratio of the dynamic feature matrix. Through the synergistic improvement of these three steps, the improved mechanism achieves a three-dimensional technical enhancement in the scenario of large-scale biomass co-firing with sidewall layered arrangement. In terms of data integrity, the gradient adaptive continuous soft membership mechanism completely recovers the 350°C to 400°C transition region data, which was discarded entirely in the first embodiment, into the segmentation result in the form of fractional weights, eliminating the negative impact of transition region information loss on the accuracy of subsequent carbon conversion rate decoupling inference. In terms of physical attribution accuracy, the cross-coupling attribution correction mechanism based on the Stefan-Boltzmann fourth-power radiation law and the spatial exponential decay kernel function ensures that the determination of the reaction zone attribution of each spatial cell in the transition zone no longer relies on empirical threshold comparisons of single-point temperature values, but rather on quantitative allocation based on the physical causal relationship of radiative energy transfer between the cell and the two reaction zones. This fundamentally solves the attribution blindness problem of the dual-source thermal radiation cross-coupling effect under the sidewall layered configuration in the first embodiment. In terms of downstream signal quality, since the entire link adopts a continuously differentiable Sigmoid function and normalized competitive allocation instead of the original hard-step discontinuity determination, there are no artificially introduced abrupt jumps from 0 to 1 in the spatially segmented combustion tensor. As a result, the numerical stability and signal-to-noise ratio of subsequent mass transfer partial derivative difference operations are systematically improved. This improvement propagates downward along the data flow link, ultimately enhancing the overall prediction accuracy and robustness of the carbon emission prediction sequence.

[0046] Accordingly, in step S22, the combustion mass transfer partial derivatives and heterogeneous time constants of the spatially segmented combustion tensor are estimated to obtain the heterogeneous time constant characteristic set. It should be understood that, in the scenario of large-scale coupled co-firing of biomass independently injected into the mixed sidewalls of large coal-fired units, the biomass feedstock fed into the furnace is rich in extremely high volatile heat-releasing components and has an extremely low ignition point. In the very short instant of contact with the high-temperature heat flow, it will generate an extremely violent deflagration oxygen-consuming effect, exhibiting pulse-like extremely rapid response characteristics. Meanwhile, the main coal powder group ground and transported by the medium-speed coal mill must go through a long physical heating tail stage from its own pre-dehydration, slow gasification and expansion to the complete oxidation and combustion of the hard coke skeleton, exhibiting an extremely slow and sluggish large hysteresis oxygen-consuming feedback trend. These two heterogeneous fuel flows, with vastly different oxidation rates and combustion heat release rates at the microscopic physical level, inevitably trigger intense asynchronous dynamic time misalignment problems within the macroscopic furnace. If static mixing mass weighing or absolute volumetric ratios are rigidly adhered to in order to mechanically assess the change in total carbon in the flue gas, this fundamental time-domain range information arising from the different rates of mass transfer and combustion will be completely missed. Therefore, in the technical solution of this application, the combustion mass transfer partial derivatives and heterogeneous time constants are further estimated on the spatially segmented combustion tensor to obtain a set of heterogeneous time constant characteristics. This allows the complex three-dimensional spatial segmentation array to be translated and compressed into a dynamic time-frequency difference quotient slope characteristic specifically characterizing the sensitivity of a particular fuel to transient oxidation and oxygen consumption changes caused by minute feed pulses at the front end. In this way, it is possible to accurately purify and lock in the unique dynamic evolution parameter imprints of biomass' rapid response and fossil coal's extreme stagnation and delay from the deep and mutually contaminated physical combustion field, providing a pure and differentiated spatiotemporal constant feature resource pool with extreme physical exclusivity and dynamic mapping analysis capabilities for subsequent deep digital decoupling networks from the foundational level.

[0047] More specifically, in a concrete example of this application, the system performs a transformation sequence regularization calculation operation from static spatial volume cutting and stripping to dynamic differential discrete column extraction. First, the system extracts and analyzes the multi-directional channel dataset with attached physical destination imprints from the spatial segmentation combustion tensor that has just been output from the upstream. It accurately strips out the subsets of real-time inference front-end absolute feed rate sequences belonging to the limits of each absolute segmentation interface and the subsets of three-dimensional local oxygen consumption decay responses corresponding to the real-time mapping, and forces them to be expanded into an array along an extremely dense synchronous clock discrete slice axis. Subsequently, the system uses the predetermined microscopic time detection section gap as the probe movement observation step length, and uses the forward approximation first-order finite difference algorithm to rigorously calculate the partial derivative slope value of the regional tiny instantaneous spatial oxygen consumption range for each type of specific fuel source to the front-end tiny feed mass step difference that induces the oxygen consumption response, thus establishing the real-time dynamic combustion mass transfer partial derivative quantity scale. Finally, the system combines the biomass mass transfer partial derivative component structure, representing the extremely rapid slope of the activation reaction, with the coal mass transfer partial derivative component structure, representing the slow and long slope of severe endothermic delay, and the calculated reaction hysteresis decay delay solidification constant vector. Following a strict time scale alignment sequence and spatial elevation plane assignment sequence, deep-level high-dimensional vector cross-linking and matrix-structured encapsulation are performed to ultimately reconstruct and distribute a heterogeneous time constant feature set that intuitively reflects the macroscopic distribution effect of the system's complete thermodynamic hysteresis evolution. The calculation process for estimating the slope of the above kinetic properties and encapsulating the heterogeneous characteristic array is as follows: Where k represents the discrete index positioning subscript used to delineate and identify the attribution characteristics of the current target calculation subject, and its value range is forcibly restricted to fall within the bio identifier group of biomass source and the coal identifier group of fossil coal source; t represents the current absolute anchor point of the benchmark synchronization moment when the central control processing core of the system implements this set of full state detection backtracking and differential derivative analysis. This represents the first-order combustion mass transfer partial derivative obtained precisely from the special simulation of target attribute fuel k at the current anchoring time point t. This indicates that the total amount of oxygen consumed is included within the target fuel k boundary spatial physical reaction zone, and the absolute extreme value of the local consumption and conversion of chemical gases is calculated at time t. This indicates the absolute and precise total weight of the target fuel k delivered into the furnace at the same anchoring time point t, after compensation and calibration to determine the corresponding burner elevation and layer section entering the furnace; This indicates the advance step size for calculating the sliding translation difference quotient of the small ranging time span pre-set when the system's main control digital clock performs full-sequence to discrete derivative calculation and range slope extraction mechanism. This represents the inner column-oriented data structure vector of the heterogeneous time constant feature set, which is fully contained and encapsulated by the dynamic response bias parameters and firmware stall delay parameters of each source after the system has been integrated and reinstalled. This indicates that the system reverse-identifies and reads the micro-scale, highly sensitive physical decay time delay constant metric parameter based on the current micro-deflagration state of biomass strong-release combustion. The large, slow, and heavy reaction decay time delay constant is a characterization parameter representing the output of the system correlation evaluation of the long-term flame retardant delay curve of the coke core.

[0048] Accordingly, in step S23, the heterogeneous time constant feature set is dimensionality-reduced and compressed to obtain the dynamic feature matrix. It should be understood that the heterogeneous time constant feature set generated after the preceding spatial physical cutting and partial derivative differential calculations not only contains highly dense parameters of the dynamic response slope distribution derived from the rapid transient response range of biomass and the long-tailed, gently sloping dynamic response of fossil coal, but also implicitly and profoundly encapsulates extremely complex redundant topological position information of the boiler heating surface's three-dimensional spatial coordinate system, as well as a large amount of background structural blank noise features. If this high-dimensional, highly spatially void-occupied, and highly redundant hashed composite heterogeneous feature vector group is directly fed into the subsequent neural network model, it will inevitably trigger an extremely severe dimensionality curse within the hidden layers, greatly increasing the computational resistance to network topology convergence and fermenting a devastating gradient explosion effect. Therefore, in the technical solution of this application, the heterogeneous time constant feature set is further dimensionality-reduced and compressed to obtain a dynamic feature matrix. This overcomes the spatial dimensional barrier by stripping away and eliminating the inefficient and redundant matrix structure occupied by massive space stations, thus reconstructing the complex high-level discrete spatiotemporal multi-source dynamic representation into a continuous real-valued reduced-rank expression format with extremely high signal-to-noise ratio and ultra-compressed pure performance. This provides a deeply denoised, highly dense, and lightweight standard driving base input matrix for subsequent dual-source transient carbon conversion rate separation and physical approximation inference performed by twin deep neural networks, completely eliminating computational burdens and ensuring the efficient and stable operation of the entire data-driven and even physical information approximation analysis chain.

[0049] More specifically, in a concrete example of this application, the system executes a rigorous cascaded compression coding pipeline that includes tensor reshaping and flattening, orthogonal projection feature rank reduction, and nonlinear boundary limiting. First, the system receives the entire set of high-dimensional heterogeneous time constant features and fully activates the underlying tensor flattening physical degradation transformation mechanism. By performing a one-dimensional dense sequence flattening operation, the multi-dimensional tensor structure, originally attached to the three-dimensional elevation space hierarchy and the deep time extension axis, is forcibly downgraded in terms of physical coordinate dimensions. For the network subsequently resolving carbon emission lag time, its core concern is merely the derivative of the rate of temperature rise and oxygen consumption determined by the intrinsic reactive energy of the fuel. The absolute spatial attachment of the fuel to which specific corner is heated is meaningless. Therefore, this flattening operation completely discards the cumbersome topological layer information forcibly derived from the thermometric grid, transforming it into a single-dimensional, coherent, and dense feature trunk. Subsequently, the system performs a deep linear inner product compression combination of the flattened and recombined one-dimensional feature array with a pre-anchored global orthogonal projection weight matrix set deep within the system kernel. This process relies on a projection matrix network that has been abstracted and shaped by principal component analysis. It forcefully excludes and eliminates the discrete and divergent redundancy associated with bidirectional fuel characteristics, retaining only the pure, differentiated core component energy that absolutely represents the mass transfer gradient of the dual-source materials. Simultaneously, a systematic bias constant matrix is ​​injected to complete numerical hedging, precisely smoothing out the long-accumulated drift and misalignment of the furnace environment background thermophysical baseline. Finally, facing the initial state of the features after the linear truncation reconstruction has converged, the system cascades and triggers a predetermined stepless smoothing nonlinear activation operator. Through boundary mapping constraints, it blocks any instantaneous outliers and high-frequency numerical cliffs that are unintentionally amplified during the micro-difference extraction process. It forces the global polarization of the features to be confined within a standard contracted dimension, ultimately solidifying, shaping, and outputting a dynamic feature matrix that is both converging and extremely accurate in terms of dynamic discrimination information and compatible with the standardized feature throughput characteristics of the deep architecture.

[0050] Specifically, in step S3, a twin deep neural network constrained by an embedded fuel ash conservation loss function is used to perform asynchronous carbon conversion rate decoupling inference on the kinetic feature matrix to obtain the transient carbon conversion rate tuple. It should be understood that in large-scale coal-fired units with a high proportion of dual-source direct coupling and co-firing, the biomass fuel injected into the sidewalls (which exhibits rapid decarbonization and burnout due to its high volatile content) and the main coal powder fed into the front and rear walls (which exhibits slow burnout with a long tail due to its extremely dense, high-fixed carbon skeleton) exhibit drastically different and completely uncoordinated asynchronous decarbonization behaviors within the complex co-firing furnace. Traditional, purely data-driven deep learning black-box models, when faced with peak-shaving stages involving large fluctuations in feed rate, cannot internalize or even adhere to the basic physical and chemical isobaric generation mechanism at a shallow level. They are prone to non-physical divergent prediction errors due to data noise, violating the law of conservation of mass, leading to frequent deviations from physical principles in the calculation of the dual-path decarbonization separation ratio. Therefore, in the technical solution of this application, a twin deep neural network with embedded fuel ash conservation loss function constraint is further used to perform asynchronous carbon conversion rate decoupling inference on the kinetic feature matrix to obtain transient carbon conversion rate tuples. In this way, parallel nonlinear orbits specially configured and designed to distinguish polarized asynchronous kinetic characteristics are used to perform independent abstract mapping and reproduction of dual-source spectral feature data. In the core loss penalty underlying architecture, the inherent non-combustible ash absolute conservation law is utilized and transformed to construct a loss function defense line for system calculation and an absolute backpropagation gradient rigid sanction suppression checkpoint against the opposite side. In this way, the probability of floating digital stacking without physical constraints can be successfully approximated and recombined into a high-confidence, hard-core deconstruction and analysis system with strong physical causal constraints. Under the premise of completely eliminating the need to rely on the extremely expensive and time-delayed online special isotope detection and isolation hardware cluster equipment at the front end of flue gas, it ensures that every component of the independent carbon release calculation ratio section generated by the system is absolutely constrained and firmly placed within the rigorous and realistic thermodynamic objective fault-tolerant extreme control base law logic. This ensures the robustness of the panoramic calculation and the generation of unbreakable second-level attribution dual-channel stripping feature data from the source system.

[0051] Figure 4 This is a flowchart of step S3 of the carbon emission prediction method for biomass direct hybrid power generation according to an embodiment of this application. Figure 4As shown, step S3 includes: S31, performing twin bi-branch forward nonlinear mapping on the kinetic characteristic matrix to obtain unconstrained conversion rate prediction data; S32, back-calculating the current transient dynamic theoretical ash residue based on the measured mass of each fuel entering the furnace and the initial burnout ratio of each branch in the unconstrained conversion rate prediction data; S33, evaluating the physical conservation deviation between the transient dynamic theoretical ash residue and the static baseline absolute total ash obtained from offline fuel industry analysis to obtain a physical conservation constraint mapping tensor; S34, performing dynamic residual calibration on the initial predicted values ​​of each branch in the unconstrained conversion rate prediction data based on the ash conservation deviation gradient in the physical conservation constraint mapping tensor to obtain the transient carbon conversion rate tuple.

[0052] Accordingly, in step S31, a twin-branch forward nonlinear mapping is performed on the kinetic feature matrix to obtain unconstrained conversion rate prediction data. It should be understood that in the scenario of large-scale coal-fired power units with a high proportion of sidewall-coupled biomass co-firing, the biomass fuel injected into the sidewalls has extremely high volatile matter and an extremely low ignition temperature, exhibiting extremely rapid low-temperature heat release pulse physical characteristics. Meanwhile, the mainstream pulverized coal fuel supplied to the front and rear walls is rich in dense, high-fixed carbon, requiring a very high ignition activation energy and undergoing an extremely long inert solid-phase heat release process at high temperatures. If a conventional single network channel is used to uniformly map and abstract these two feature sequences, which exhibit extremely polarized and heterogeneous characteristics at the kinetic level, the network's bottom-level gradient will inevitably fall into severe feature mixing interference and weight update oscillations, making it impossible to simultaneously characterize the distinctly different fast and slow reaction evolution separation curves. Therefore, in the technical solution of this application, a twin-branch forward nonlinear mapping is further performed on the kinetic feature matrix to obtain unconstrained conversion rate prediction data. This constructs two physically parallel and differently configured deep forward computation paths, forcing the heterogeneous ultra-fast response and extremely delayed kinetic features to complete cross-domain parallel analysis from physics to digital within their respective independent parameter spaces. In this way, the cross-source pollution and feature entrainment caused by the high volatile biomass anomaly data on the long-tail coking characteristics of pulverized coal can be completely cut off. Before the subsequent rigid physical quality loss reconstruction is introduced, the high-fidelity approximation extraction of dual-source asynchronous physical combustion attributes is completed first within the digital hidden layer space. This provides a highly exclusive and discriminative initial state deduction base combination platform that is free from chaotic peak disturbances for the final absolute ash conservation residual verification of the overall prediction system.

[0053] Specifically, in this embodiment, a twin-branch forward nonlinear mapping is performed on the kinetic feature matrix to obtain unconstrained conversion rate prediction data. This includes: feeding the kinetic feature matrix in parallel into two decoupled mapping branches of a twin deep neural network; the first branch uses dedicated neuron weights adjusted for high volatile matter and low-temperature heat release characteristics to output the initial oxidation burnout ratio of biomass via nonlinear forward propagation; the second branch uses a parallel weight structure adjusted for high fixed carbon and high-temperature inert heat release characteristics to output the initial oxidation burnout ratio of coal via nonlinear forward propagation; and the outputs of the two branches are constrained to a probability range of zero to one and combined and encapsulated using a normalized activation function to obtain unconstrained conversion rate prediction data.

[0054] More specifically, in a concrete example of this application, the system executes a nonlinear fitting pipeline that maps a high-dimensional dynamic representation to a dual-path quasi-probability space. First, the system acquires the compressed and denoised dynamic feature matrix and synchronously feeds it into a pre-defined Siamese deep neural network system in a parallel bus mode. This system is internally anchored by physical laws and divided into two logically isolated and topologically parallel decoupled mapping branches. The first branch utilizes a dedicated neuron weight matrix adjusted for high volatile matter and low-temperature heat release characteristics to accurately pick up high-frequency energy pulsation signals caused by biomass deflagration in the dynamic features during nonlinear forward propagation, thereby outputting the initial oxidation and burnout ratio of the biomass. Simultaneously, the system drives the second branch to load a pre-defined parallel weight structure for high fixed carbon and high-temperature inert heat release characteristics. This structure is specifically responsible for detecting the characteristic accumulation process of delayed combustion of pulverized coal in the high-temperature swirling core region, thereby outputting the initial oxidation and burnout ratio of the coal. Since the above mapping process is merely a mathematical approximation driven by historical data and has not yet introduced hard physical constraints such as mass conservation, the system will finally load a normalized activation function at the end of the dual-path fusion to perform a nonlinear limiting operation. This forces the output results to converge to a legal probability interval of zero to one for combination and encapsulation, ultimately generating unconstrained conversion rate prediction data with clear source resolution significance. When performing the above source-specific abstract mapping for heterogeneous fuel characteristics, the core logic within the system is implemented through the following nonlinear regression operator formula: Where t represents the absolute reference time anchor point for the synchronous inference calculation of the current execution system; k represents the branch location index identifier representing the intrinsic characteristics of different fuels, and its value is strictly limited to representing biological origin attributes. and the properties representing fossil coal ; This represents the preliminary predicted proportion of oxidation burnout, calculated independently by the system twin branch at time t and not yet verified by physical mass balance. This represents a nonlinear logistic regression function with globally continuous differentiability, used to perform normalization suppression in the interval from zero to one. This represents the weight matrix for extracting specific features, which is obtained in advance during the model building phase by training on independent combustion samples of biomass or coal. The dynamic feature matrix representing the current input system, carrying information about the slope caused by the rate of combustion; This represents the bias constant vector used for systematic deviation correction during the branch mapping process. The index term k in the above equation divides the hybrid dynamic characteristic matrix into two independent evolution trajectories, which is physically equivalent to establishing a virtual combustion process monitor for the biomass jet and the pulverized coal particle group within the furnace. Weighting operator The system undertakes the core identification function. The weights of the biomass branch are extremely sensitive to low-gradient temperature rise responses, enabling them to extract the oxygen-consuming pulses caused by the instantaneous combustion of high-volatile materials such as straw from complex sensor fluctuations. Meanwhile, the weights of the coal branch capture the long-tail features of coke oxidation that are masked by rapid combustion signals through neural connections with large receptive fields. Through this twin-like forward nonlinear mapping, the system successfully overcomes the limitations of black-box models in handling the homogenization of heterogeneous fuels, ensuring that the subsequent unconstrained conversion rate prediction data includes the independent physical evolution depth information of each of the two fuel sources. This completely solves the problem of carbon emission source tracking in the context of overlapping thermal fields caused by the layered arrangement of the sidewalls.

[0055] Accordingly, in step S32, the current transient dynamic theoretical ash residue is inferred from the measured mass of each fuel entering the furnace and the initial burnout ratio of each branch in the unconstrained conversion rate prediction data. It should be understood that, in the actual physical process of biomass direct-mixed power generation, the decarbonization process after fuel enters the furnace is an invisible chemical reaction. Unconstrained conversion rate prediction data generated solely based on data is essentially a probabilistic estimate lacking material support, and is prone to outputting values ​​that violate thermodynamic principles when the load fluctuates drastically or the feed is unstable. To bring this blind numerical inference back to the track of physical reality, a conserved quantity that does not change with combustion fluctuations must be introduced as a verification benchmark. Therefore, in the technical solution of this application, the current transient dynamic theoretical ash residue is further inferred from the measured mass of each fuel entering the furnace and the initial burnout ratio of each branch in the unconstrained conversion rate prediction data, thereby establishing a hard mathematical correlation between the front-end material feed and the back-end oxidation result using the law of conservation of mass. In this way, the intangible gaseous transformation process can be mapped into the tangible solid matter retention logic, providing a computational intermediate with an absolute causal chain for subsequent deviation assessment and dynamic correction in the physical dimension.

[0056] More specifically, in a specific example of this application, the system performs a backtracking conversion operation from known input phenological quantities to theoretical reaction margins. First, the system retrieves in real-time the measured input mass of each fuel from the coal feeders and biomass conveying devices at each synchronous moment. This data represents the discrete physical mass of the material currently input into the thermodynamic reaction cycle. Then, the system extracts the unconstrained conversion rate prediction data produced by the aforementioned twin network, and uses the initial oxidation and burnout ratios of biomass and coal to extract the corresponding percentage of unburned residue based on the subtraction conservation logic. Next, the system multiplies the measured input mass of different fuel types with their corresponding percentage of unburned residue, thereby deconstructing and synthesizing the transient dynamic theoretical ash residue composed of both biomass and coal sources. When performing the above-mentioned quantitative monitoring of the internal mass balance state of the furnace, the system defines the following backtracking calculation equation: Where t represents the absolute time anchor point for the synchronous sampling and back-calculation of the execution system; k represents the classification cursor used to distinguish the physical properties of the fuel; This represents the transient dynamic theoretical ash residue generated by the system logic layer at time t, in kilograms per second. This represents the measured mass flow rate of fuel k entering the furnace at time t, provided in real time by the front-end metering equipment, in kilograms per second. This represents the initial oxidation burnout ratio scalar of fuel k at that moment, derived from the system's twin branch. Combined with the mass balance engineering scenario of sidewall layered coupled co-firing, the above inverse equation, through structured synthesis of physical quantities, accurately anchors the physical centroid for each bit of model prediction data. The terms in the equation... In physical semantics, this represents the theoretical proportion of residue in fuel k that has not yet escaped as gaseous carbon dioxide and still resides in the solid phase of the furnace; this is achieved by comparing it with the actual physical inflow rate fed into the furnace. By establishing multiplicative correlations, the system can use measured inflow mass to constrain virtual reaction trends. The engineering significance of this processing logic lies in its transformation of the problem of the rate of a chemical reaction, which is difficult to perceive directly, into a problem of assessing solid residues that follows the law of conservation of mass. This allows the system to monitor in real time the digital illusions of matter disappearing out of thin air or illegal energy surges caused by neural network fitting biases, thus laying a solid foundation for establishing a carbon decoupling prediction system with physical and logical self-consistency based on the principle of matter conservation.

[0057] Accordingly, in step S33, a physical conservation deviation assessment is performed on the transient dynamic theoretical ash residue and the static baseline absolute total ash obtained from offline fuel industry analysis to obtain a physical conservation constraint mapping tensor. It should be understood that, because the coupled combustion process of biomass and pulverized coal is strictly governed by the objective physical law of conservation of mass, although neural networks can fit and provide a preliminary burnout ratio based on the fluctuations of previously acquired signals, this reasoning process based on discrete numerical patterns often has uncontrollable blindness, easily leading to significant logical conflicts between the theoretical residual mass calculated and the inherent mineral composition ratio of the fuel, which does not change with operating conditions. Without introducing a hard physicochemical benchmark obtained from pre-analysis of the fuel for collision verification, the system will be unable to detect the mass non-conservation error in the model output caused by data noise interference or nonlinear abrupt changes, thus losing the dynamic constraint means for the physical compliance of carbon emission prediction results. Therefore, in the technical solution of this application, a physical conservation deviation assessment is further performed on the transient dynamic theoretical ash residue and the static baseline absolute total ash obtained from offline measurement in fuel industry analysis to obtain a physical conservation constraint mapping tensor. This transforms the rigid boundary of stoichiometry into a quantized deviation gradient with differentiable properties. This allows for precise identification of the severity of imbalance in the physical reality of the neural network output value, providing a set of correction instructions with an absolute causal chain for subsequent dynamic residual calibration driven by the model.

[0058] More specifically, in a concrete example of this application, the system executes a closed-loop evaluation sequence from static quality benchmark to dynamic deviation tensor encapsulation. First, the system fully loads the transient dynamic theoretical ash residue synthesized by the pre-calculation step through instantaneous feed mass and predicted ratio. Simultaneously, the system retrieves the biomass ash ratio and pulverized coal ash ratio, pre-determined through offline measurements of the corresponding batch of fuel industry analysis from the configuration management database, and combines them with the instantaneous feed flow rate fed back by the feed sensor to synthesize the static benchmark absolute total ash flow rate that the system must adhere to and cannot exceed at the physical reality level within this time segment. Subsequently, the system performs a L2 norm squared difference collision operation on the theoretical value representing the fitted state of the numerical inference and the benchmark value representing the underlying physical reality. By quantifying the deviation distance of the two types of physical quantities in Euclidean space, a loss evaluation function is constructed. Finally, the system performs a high-dimensional mapping on the real-time loss scalar obtained from the evaluation, along with the spatiotemporal characteristics of the multi-source feed, and finally solidifies and encapsulates it into a physical conservation constraint mapping tensor that guides the model to approximate the real physical state. In performing the above quantitative assessment of the deviation from the law of conservation of mass, the system defined and ran the following deviation assessment loss equation, which includes physical prior constraints: Where t represents the system synchronization time anchor point for performing this discrete evaluation calculation; k represents the classification index used to separate fuel attributes from different sources; This represents the total feedback loss scalar of the system that violates the law of conservation of physical mass, calculated at time t. It constitutes the core guiding numerical component inside the physical conservation constraint mapping tensor. It represents the transient dynamic theoretical ash residue formed under the preliminary estimation state of the system twin network, which has a digital illusion-like quality; This represents the instantaneous total mass flow rate of fuel k entering the furnace at time t, as measured by the feed probe at the physical front end. This represents the percentage by mass of non-combustible absolute ash in the base composition of fuel k, which is pre-calibrated through industrial testing and analysis and exists as a physical hard contract. The summation term within the equation. It fulfills the core function of truth comparison, disregarding any algorithmic reasoning and directly utilizing measured flow rates. By locking in the percentage of dead mass that must be carried into the furnace, the static physical ceiling for ash production at the bottom of the furnace at this moment is defined. In contrast, This reflects the current wishful thinking of AI neural networks regarding the intensity of combustion. Through the L2-norm squared operation applied to the outer layer, the system can capture in real-time and acutely the unreliable behavior of theoretical estimates leading to illegal ash annihilation or the unauthorized generation of minerals due to overfitting or signal lag. The larger this difference, the stronger the deviation gradient potential energy output to the physical conservation constraint mapping tensor. The system thus obtains a forced traction computing field capable of forcibly pulling back to the physical baseline, eliminating digital divergence, and ensuring that the dual-source carbon decoupling logic always operates within the strict boundaries of stoichiometry by adjusting the burnout ratio.

[0059] Accordingly, in step S34, based on the ash conservation deviation gradient in the physical conservation constraint mapping tensor, dynamic residual calibration is performed on the initial predicted values ​​of each branch in the unconstrained conversion rate prediction data to obtain the transient carbon conversion rate tuple. It should be understood that, in the complex boiler co-firing reaction field, the black-box deep learning branch only captures the decarbonization law at the representational level through numerical fitting, lacking the ability to anchor the mass at the thermodynamic level. This leads to the unconstrained conversion rate prediction data generated purely by data driving being highly prone to deviations that violate the iron law of conservation of inorganic matter under transient load fluctuations. This physical-level failure directly undermines the logical foundation of decoupled prediction, causing the subsequently calculated fossil carbon and biocarbon emissions to lose their industrial logical consistency. Therefore, in the technical solution of this application, based on the ash conservation deviation gradient in the physical conservation constraint mapping tensor, dynamic residual calibration is performed on the initial predicted values ​​of each branch in the unconstrained conversion rate prediction data to obtain transient carbon conversion rate tuples. This allows for the transformation of the hard constraints of stoichiometry into self-feedback correction mechanisms for each fuel branch prediction value through an online reverse compensation mechanism that forcibly embeds physical rules at the inference process level. This completely suppresses the non-physical illusion oscillations generated by the neural network, correcting the virtual probabilistic inference into accurate analytical results supported by the laws of material entities, ensuring that each set of transient carbon conversion rate sequences output by the system operates strictly within the real physical boundaries defined by thermodynamic laws.

[0060] More specifically, in a concrete example of this application, the system executes an error correction pipeline that includes gradient-guided hedging and transient feature alignment encapsulation. First, the system analytically extracts the ash conservation deviation gradient from the input physical conservation constraint mapping tensor. This gradient component quantifies the degree of deviation of the theoretical residual ash from the static physicochemical benchmark caused by overfitting or underfitting in the current model and guides the direction of correction. Subsequently, the system extracts the initial predicted values ​​of the corresponding biomass and coal branches from the unconstrained conversion rate prediction data and drives the residual calibration logic. Under this logic, the system scales the ash conservation deviation gradient using a preset sensitivity step size, generating a compensation component opposite to the direction of the physical deviation, and directly injects it into the output mapping layer of the Siamese network, performing a single-step dynamic hedging operation to reduce invalid numerical fluctuations that violate the law of ash mass conservation. Finally, the system aligns the coal carbon conversion rate feature scalar, which has been forcibly calibrated by the physical iron law, with the biomass carbon conversion rate feature scalar, encapsulating and producing a transient carbon conversion rate tuple with physical certainty. In the process of performing dynamic correction through gray conservation gradient and achieving physical logic closure, the system deploys the following first-order residual correction mathematical model equation: Where t represents the reference time anchor point for performing this discrete synchronous sampling and dynamic calibration; k represents the classification attribute identifier describing the fuel combustion separation characteristics; This represents the final burnout ratio output value of the k-th type of fuel, calibrated according to physical rules, contained in the final generated transient carbon conversion rate tuple; This represents the initial predicted value of the k-th branch generated in advance in the corresponding unconstrained conversion rate prediction data; This represents the dynamic residual calibration step size coefficient, which is a built-in system definition used to control the guiding force of physical boundary constraints. This represents the ash conservation deviation gradient component extracted from the physical conservation constraint mapping tensor, reflecting the sensitivity of the total physical deviation to the predicted value of a specific branch. Combined with the mass balance engineering scenario of sidewall layered coupled co-firing, the above calibration operator equation endows the AI ​​prediction model with extremely robust self-correcting physical properties. The derivative term within the equation... It acts as the "physical enforcer" within the system. When the model assigns an excessively high initial burnout ratio to a certain fuel (such as a high-volatile material), resulting in a calculated theoretical ash content significantly lower than the absolute total ash content brought in by the actual feed, this gradient term immediately exhibits a large positive correction potential. This is achieved by comparing it with the step size coefficient. By multiplying and then performing a subtraction operator mapping, the system forcibly adjusts the initial value while preserving the data-driven fitting slope. The non-physical portion that is artificially inflated due to numerical fluctuations is deducted. The physical significance of this calculation sequence lies in the fact that, through microscopic control of the branches of each of the two source paths, it ensures that, regardless of fluctuations in the external feed, the two sets of scalars produced must perfectly reproduce the mass equilibrium state of the minerals fed into the boiler after merging and backtesting. This completely eliminates the persistent problem of unreliable and discrete carbon emission prediction data caused by the illusion of black-box algorithms, and lays the foundation for the subsequent separate stoichiometric decoupling calculations with a transient carbon conversion rate tuple substrate with absolute physical reliability.

[0061] Specifically, in step S4, based on the transient carbon conversion rate tuple and the ratio of hysteresis compensation feed and carbon content stored in the aligned combustion state tensor, stoichiometric decoupling calculations are performed on the biomass source and the coal source to obtain decoupled carbon emission data. It should be understood that in large-scale biomass co-firing projects with layered sidewalls, biomass and coal, two heterogeneous fuels, not only have adjacent spatial distributions but also exhibit extreme asynchrony in their energy evolution. Specifically, biomass fuel, due to its extremely high volatile matter content, exhibits second-level deflagration, with minimal thermal lag in other reactions, while pulverized coal particles require a much longer time span to complete the entire process from heating and dehydration to coke burnout, exhibiting significant thermal lag characteristics. This means that the transient emission signal detected in the economizer tail flue is essentially the cumulative reaction result of fuels fed into the furnace at different historical moments at the current node. If the current feed rate is used for static calculation, it will inevitably lead to a severe phase misalignment between the carbon emission calculation data and physical reality. Therefore, in the technical solution of this application, based on the transient carbon conversion rate tuple and the ratio of hysteresis compensation feed and carbon content stored in the aligned combustion state tensor, stoichiometric decoupling calculations are performed on biomass and coal sources respectively to obtain decoupled carbon emission data. This is used to correct the phase difference between fuel delivery and combustion reaction using a kinetic time delay compensation mechanism, and the abstract conversion ratio is reduced to a concrete mass of released substances according to the law of chemical conservation. In this way, the scattered real-valued flows of fossil carbon and biomass carbon can be completely separated, providing a decoupled original data body with chemical causal reliability for constructing a legal, compliant, and time-evolution-accurate carbon footprint tracking sequence.

[0062] More specifically, in the embodiments of this application, step S4 includes: based on the thermodynamic lag parameter determined by the inherent combustion kinetic time difference of the two fuels, tracking and extracting the historical delayed feed amounts of coal and biomass in the aligned combustion state tensor based on propagation lag to obtain a thermodynamic delayed feed matrix; aligning the historical delayed feed amounts of coal and biomass in the thermodynamic delayed feed matrix, the corresponding coal carbon conversion rate feature scalar and biomass carbon conversion rate feature scalar in the transient carbon conversion rate tuple, and the carbon element content ratio obtained by pre-offline analysis of each fuel by vector splicing according to fuel category to obtain a dual-source calculation preparation feature set; based on the inherent constant ratio relationship between the atomic weight of carbon element and the molecular weight of carbon dioxide, performing molar mass ratio chemical equivalent conversion and decoupling identification assembly on the combination vector of biomass source and the combination vector of coal source in the dual-source calculation preparation feature set to obtain decoupled carbon emission data.

[0063] Specifically, in a concrete example of this application, firstly, the system parses and retrieves the thermodynamic lag parameters determined by the inherent combustion kinetic time differences of the two fuels, and enters the historical storage stack of the aligned combustion state tensor. It then tracks and extracts the historical delayed feed amounts of coal and biomass based on propagation lag to obtain a thermodynamic delayed feed matrix. Subsequently, the system performs structured alignment of multidimensional features, aligning the historical delayed feed amounts of coal and biomass in the thermodynamic delayed feed matrix, the corresponding coal carbon conversion rate feature scalars and biomass carbon conversion rate feature scalars in the transient carbon conversion rate tuples, and the carbon element content ratio of each fuel as determined by pre-experimental analysis. This is done by performing high-dimensional vector splicing and alignment operations according to the fuel category, constructing a dual-source computational preparation feature set containing all antecedent components. Finally, the system loads a stoichiometric evaluation operator. Based on the inherent constant ratio between the atomic weight of carbon and the molecular weight of carbon dioxide, it performs molar mass ratio stoichiometric conversion on the combination vectors of biomass sources and coal sources in the dual-source calculation preparation feature set, and performs decoupling identification assembly on the metadata tags of biosource attribution and fossil source attribution superimposed on the calculation results. Finally, it outputs decoupled carbon emission data with second-level analytical resolution.

[0064] When performing the transient decoupling calculation for carbon release from heterogeneous fuels, the system internally deploys the following backtracking extraction formula and its corresponding stoichiometric equivalent conversion equation set: The physical logic equation used to perform historical feed rate tracking extraction is as follows: Where t represents the baseline sampling synchronization time anchor point for this carbon accounting; k represents the classification code subscript of the fuel source, corresponding to the bio and coal attributes respectively; This represents the historical delayed feed amount of the k-th type of fuel that contributes the most to the current emission results, retrieved at the current time t. It constitutes the core term of the thermodynamic delayed feed matrix. Represents the alignment combustion state tensor The time-dimension slice extraction function is implemented; This represents the pre-calibrated thermodynamic hysteresis parameter for fuel k, which is the time constant for the material to move from the feed port to the effective reaction zone.

[0065] The stoichiometric equations for performing the decoupled carbon dioxide mass production calculations from dispersed sources are as follows: in, and ) represent the fossil-source carbon dioxide release mass flow rate and the biological-source carbon dioxide release mass flow rate contained in the decoupled carbon emission data output by the system at time t, respectively, in kilograms per second; and This represents the constant ratio of the corresponding dry-basis carbon element content of the fuel, obtained from the physicochemical analysis of the fuel. and 44 and 12 represent the real-time carbon conversion rate characteristic scalars extracted from the transient carbon conversion rate tuple, respectively; 44 and 12 represent the molecular weight of carbon dioxide and the atomic mass constant of carbon atoms, respectively.

[0066] The above set of equations, through the nesting of physical time delays and chemical conservation, profoundly quantifies the causal chain of carbon footprint within complex co-firing fields. The equations contain... This accurately reflects the impact of flow field dynamics on material transport, ensuring that the model uses past causes to explain present effects, thus eliminating transient false carbon emission errors caused by feed fluctuations due to peak shaving. The molar mass ratio coefficient in the core formula... It performs the function of mass conversion amplification, multiplying the amount of single-atom carbon carried by the fuel by an oxidation ratio that has been highly corrected by the model. Ultimately, this is transformed into the physical carbon dioxide emissions value that directly determines the trading amount. Through this decoupled label assembly method of separate measurement, the system successfully separates the total flue gas flow mixed at the chimney outlet into two digital asset flows with completely different environmental attributes. This allows the decoupled carbon emission data to clearly outline the real-time carbon emission reduction contribution of biomass co-firing, greatly enhancing the strength of physical evidence for the predicted sequence in responding to carbon compliance verification.

[0067] Specifically, in step S5, a temporal convolutional sequence network with an attention mechanism is used to perform trend evolution and forward prediction on the decoupled carbon emission data to obtain a carbon emission prediction sequence. It should be understood that during the dynamic operation of large coal-fired units performing deep peak shaving and high-proportion biomass co-firing, carbon emission processes are not isolated point-in-time data, but rather exhibit a highly correlated evolutionary trend influenced by load command control, fuel switching disturbances, and nonlinear hysteresis of the thermodynamic cycle. Simple decoupled real-time values ​​can only reflect the historical and current instantaneous states, failing to provide a basis for forward-looking decisions regarding carbon trading quota settlement. Furthermore, traditional linear prediction models struggle to capture the high-frequency oscillation slope characteristics caused by biomass feed fluctuations or pulverizer start-up and shutdown, easily leading to digital illusions that deviate from the physical limits of boiler flue gas. Therefore, in the technical solution of this application, a temporal convolutional sequence network with an attention mechanism is further used to perform trend evolution and forward prediction on the decoupled carbon emission data to obtain a carbon emission prediction sequence. This utilizes a deep temporal sensing architecture to mine the inertial patterns of fossil carbon flow and introduces physical constraints on total flue gas volume to perform end-point residual trimming. In this way, we can successfully elevate pure retrospective statistics to a forward-looking management level with physical boundary guarantees, ensuring that the generated emission trends can not only respond sensitively to drastic changes in unit combustion conditions, but also absolutely meet the gaseous physical limits of the boiler tail flue gas system, providing a high-confidence logical spectrum for the refined management of carbon quotas.

[0068] More specifically, in this embodiment, step S5 includes: extracting historical evolution curves of fossil carbon emission portions from decoupled carbon emission data for several consecutive natural cycles over a predetermined period to construct a sliding observation window; calculating the transient fossil carbon flow oscillation slope between adjacent time slices within the sliding observation window, and combining the original carbon release mass sequence and the oscillation slope along the feature dimension to obtain a historical evolution tensor; performing forward inference on the historical evolution tensor based on an attention-based temporal convolution system to obtain a primary forward prediction sequence; and performing physical confidence verification on the predicted values ​​of each future time slice node in the primary forward prediction sequence based on the exhaust oxygen concentration indicated by the tail probe and the total exhaust volume rate in the aligned combustion state tensor to obtain a carbon emission prediction sequence.

[0069] Specifically, in a concrete example of this application, the system first extracts the fossil-source carbon emission portion from the real-time generated decoupled carbon emission data, extracts the historical evolution curves for several consecutive preset natural cycles, and loads them into a shift register to construct a sliding observation window. Subsequently, the system performs a first-order discrete difference operation on each adjacent time slice within the window to calculate the transient fossil carbon flow oscillation slope, thereby extracting the acceleration characteristics of emission changes with operating conditions. The system then combines the original carbon release mass sequence and the oscillation slope along the feature dimension using channel stacking to obtain a historical evolution tensor. Next, the system feeds the historical evolution tensor into a preset temporal convolutional sequence network containing a self-attention mechanism and dilated convolutional layers. By assigning high-attention weights to inflection points, it performs forward-looking inference across time periods to obtain a preliminary forward-looking prediction sequence. Finally, to prevent logical drift caused by the independent operation of the algorithm model, the system analyzes and retrieves the exhaust oxygen concentration and total exhaust volume rate indicated in real time by the tail probe in the aligned combustion state tensor. In this way, the physical emission limit threshold of carbon release under the current ventilation conditions is determined. Each prediction node in the primary look-ahead prediction sequence is then checked for physical confidence and the residual is clipped using this threshold. Finally, a carbon emission prediction sequence that reflects the physical reality is solidified and generated.

[0070] When performing the aforementioned high-dimensional predictive evolution of future emission trends, the system internally deployed and ran the following mathematical computational model, which includes slope abstraction, look-ahead inference, and physical constraint verification: The discrete mapping equation used to extract historical spatiotemporal fluctuation features and reconstruct the feature tensor is: Where t represents the current synchronization time anchor point for performing predictive sampling; m represents the step size of historical backtracking. Indicates time The corresponding real-time release of fossil sources in the decoupled carbon emission data; This indicates the time point calculated and generated by the system ( The corresponding transient fossil carbon flow oscillation slope scalar; This represents the final synthesized output and the historical evolution tensor used to drive network inference. This system of equations combines the two-dimensional features of displacement value and rate of change into one through a concat cascade operation, ensuring that the carbon flow drastic characteristics caused by sudden changes in unit load are enhanced before entering the network.

[0071] The material balance equation for determining the physical emission ceiling based on the upper limit of the exhaust volume of the induced draft chimney is as follows: in, This represents the highest instantaneous physical emission limit threshold for carbon emissions generated by the system in real-time dynamic inversion, expressed in kilograms per second. This represents the total volumetric flow rate of the exhaust gas at the economizer tail under standard conditions provided by the aligned combustion state tensor; This indicates the base percentage of oxygen concentration in the exhaust gas; This represents the density constant of carbon dioxide under standard conditions. This equation reveals the aerodynamic red line of the combustion reaction, namely, that for a given total exhaust volume, the more oxygen consumed, the higher the carbon emissions, and the corresponding total oxygen consumption limit constitutes an insurmountable physical confidence upper limit.

[0072] The physical confidence constraint operator formula for performing final residual verification and clipping is as follows: Where j represents the forward projection step number; This represents the value predicted at step j in the primary look-ahead prediction sequence; This indicates the finalized values ​​that, after verification, belong to the carbon emission prediction sequence; This represents the penalty step size weighting factor used for smoothing clipping.

[0073] It is worth mentioning that, considering the engineering application scenario of layered sidewall layout accompanied by frequent loading and unloading of medium-speed coal mills, the above mathematical calculation system demonstrates a strong anti-illusion prediction capability. The equations in... It captures the advanced power response during the instantaneous acceleration of the feed belt frequency converter, enabling the network with an attention mechanism to anticipate the impending carbon surge due to the dramatic increase in fuel quantity. And the most crucial... The physical verification logic essentially provides a layer of physical protection for the neural network. It completely blocks a scenario where an industrial system might falsely predict a massive carbon emission peak far exceeding the rated airflow of the induced draft fan due to training sample noise. This operator immediately identifies this violation of basic flue gas volume balance and... The function performs a forced flattening action. This ensures that the carbon emission forecast sequence not only has statistically significant trend accuracy on the time axis, but also fully conforms to the legitimate data flow of the actual throughput capacity of the 600MW supercritical boiler on a physical scale, providing rigid physical logic support for the risk-controlled management of carbon assets.

[0074] In summary, the carbon emission prediction method for biomass direct-hybrid power generation according to the embodiments of this application is explained. First, it utilizes aerodynamics of the flow field to compensate for time lag and achieve mirror alignment of multi-source sensor data. Then, through three-dimensional temperature field reaction zone identification technology, the asynchronous ignition signal is converted into dynamic characteristics, and coupled with a twin deep neural network with built-in ash mass conservation constraints, achieving white-box decoupling of the transient carbon conversion rates of biomass and coal. Finally, based on this, combined with stoichiometric conversion and physical emission limit verification, a forward-looking trend prediction with physical causal terms is completed. This concept fundamentally solves the measurement disconnect caused by asynchronous combustion lag, achieving second-level accurate separation and compliant prediction of dual-source carbon emissions without relying on expensive hardware.

[0075] Furthermore, a carbon emission prediction system for biomass direct hybrid power generation is also provided.

[0076] Figure 5 This is a block diagram of a carbon emission prediction system for biomass direct hybrid power generation according to an embodiment of this application. Figure 5 As shown, the carbon emission prediction system 100 for biomass direct-hybrid power generation according to an embodiment of this application includes: a data spatiotemporal frequency alignment module 110, used to perform spatiotemporal frequency alignment processing on the acquired raw sensor data stream based on the flow field aerodynamic delay transmission module to obtain an aligned combustion state tensor; a dynamic feature extraction module 120, used to perform reaction zone identification and asynchronous thermodynamic feature extraction on the aligned combustion state tensor to obtain a dynamic feature matrix; and a carbon conversion rate decoupling inference module 130, used to perform twin deep... A degree neural network is used to perform asynchronous carbon conversion rate decoupling inference on the kinetic feature matrix to obtain transient carbon conversion rate tuples; a stoichiometric decoupling conversion module 140 is used to perform stoichiometric decoupling conversion on biomass sources and coal sources respectively based on the transient carbon conversion rate tuples and the ratio of hysteresis compensation feed and carbon element content stored in the aligned combustion state tensor to obtain decoupled carbon emission data; a trend evolution prediction module 150 is used to perform trend evolution and forward prediction on decoupled carbon emission data through a temporal convolutional sequence network with attention mechanism to obtain carbon emission prediction sequences.

[0077] As described above, the carbon emission prediction system 100 for biomass direct hybrid power generation according to the embodiments of this application can be implemented in various wireless terminals, such as servers with carbon emission prediction algorithms for biomass direct hybrid power generation. In one possible implementation, the carbon emission prediction system 100 for biomass direct hybrid power generation according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the carbon emission prediction system 100 for biomass direct hybrid power generation can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the carbon emission prediction system 100 for biomass direct hybrid power generation can also be one of many hardware modules of the wireless terminal.

[0078] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for predicting carbon emissions from a biomass direct-mix coupled power generation, characterized by, include: S1, based on the flow field aerodynamic delay transmission module, performs spatiotemporal frequency alignment processing on the acquired raw sensor data stream to obtain the aligned combustion state tensor; S2, perform reaction zone identification and asynchronous thermodynamic feature extraction on the aligned combustion state tensor to obtain the dynamic feature matrix; S3, based on the Siamese deep neural network constrained by the embedded fuel ash conservation loss function, performs asynchronous carbon conversion rate decoupling inference on the kinetic feature matrix to obtain the transient carbon conversion rate tuple; S4. Based on the transient carbon conversion rate tuple and the ratio of hysteresis compensation feed and carbon content stored in the aligned combustion state tensor, stoichiometric decoupling calculations are performed on biomass sources and coal sources respectively to obtain decoupled carbon emission data. S5 uses a temporal convolutional sequence network with an attention mechanism to perform trend evolution and forward prediction on decoupled carbon emission data to obtain a carbon emission prediction sequence.

2. The method of claim 1, wherein the carbon emission prediction of biomass direct co-firing power generation is characterized by, The raw sensor data stream includes the real-time transient inflow rate of the coal feeders and biomass conveying devices at each level of the furnace, the primary air velocity and secondary air velocity, the three-dimensional temperature cross-section matrix of each level of the burner fed back by the acoustic temperature measurement array, and the exhaust oxygen concentration at the tail of the economizer.

3. The carbon emission prediction method for biomass direct-hybrid power generation according to claim 1, characterized in that, Step S1 includes: The pulse rejection of real-time transient furnace inlet flow rate, primary air velocity and secondary air velocity is performed based on the median filter, and the bad zone reconstruction processing of isolated invalid measurement points in the three-dimensional temperature cross section matrix is ​​performed based on the filter matrix to obtain the preprocessed sensor data stream. Within a sliding window, multi-source heterogeneous signals with inconsistent sampling frequencies in the preprocessed sensor data stream are resampled and mapped to the same frequency to obtain a time-frequency data matrix. Using the total spatial transmission delay as a compensation, the front-end feeding data, three-dimensional temperature field data, and tail-end exhaust oxygen concentration data in the time-frequency data matrix are remapped along the time axis to obtain the aligned combustion state tensor.

4. The carbon emission prediction method for biomass direct-hybrid power generation according to claim 1, characterized in that, Step S2 includes: The spatially segmented combustion tensor is obtained by asynchronously dividing the spatial reaction zone of the ignition temperature field in the three-dimensional temperature cross-section data of the aligned combustion state tensor. The combustion mass transfer partial derivatives and heterogeneous time constants of the spatially segmented combustion tensor are estimated to obtain the characteristic set of heterogeneous time constants. The heterogeneous time constant feature set is dimensionality reduced and compressed to obtain the dynamic feature matrix.

5. The carbon emission prediction method for biomass direct-hybrid power generation according to claim 1, characterized in that, Step S3 includes: A twin-branch forward nonlinear mapping is performed on the dynamic characteristic matrix to obtain unconstrained conversion rate prediction data; Based on the measured mass of each fuel entering the furnace and the preliminary burnout ratio of each branch in the unconstrained conversion rate prediction data, the current transient dynamic theoretical ash residue is inferred. Physical conservation deviation assessment was performed on the transient dynamic theoretical ash residue and the static baseline absolute total ash obtained from offline measurement in fuel industry analysis to obtain the physical conservation constraint mapping tensor. Based on the ash conservation bias gradient in the physical conservation constraint mapping tensor, dynamic residual calibration is performed on the initial predicted values ​​of each branch in the unconstrained conversion rate prediction data to obtain the transient carbon conversion rate tuple.

6. The carbon emission prediction method for biomass direct hybrid power generation according to claim 5, characterized in that, A twin-branch forward nonlinear mapping is performed on the dynamic characteristic matrix to obtain unconstrained conversion rate prediction data, including: The dynamic feature matrix is ​​fed in parallel into two decoupled mapping branches of the twin deep neural network; The first branch utilizes specialized neuronal weights adjusted for high volatile matter and low-temperature heat release characteristics to output the initial oxidation and burnout ratio of biomass via nonlinear forward propagation; The second branch utilizes a parallel weighted structure adjusted for high fixed carbon and high-temperature inert heat release characteristics to output the initial oxidation and burnout ratio of coal via nonlinear forward propagation. By using a normalized activation function, the outputs of the two branches are constrained to a probability range of zero to one and combined and encapsulated to obtain unconstrained conversion rate prediction data.

7. The carbon emission prediction method for biomass direct-hybrid power generation according to claim 1, characterized in that, Step S4 includes: Based on the thermodynamic lag parameter determined by the inherent combustion kinetic time difference of the two fuels, the historical delayed feed amounts of coal and biomass in the aligned combustion state tensor are tracked and extracted based on the propagation lag to obtain the thermodynamic delayed feed matrix. The historical delayed feed amounts of coal and biomass in the thermodynamic delayed feed matrix, the corresponding coal carbon conversion rate feature scalars and biomass carbon conversion rate feature scalars in the transient carbon conversion rate tuple, and the carbon element content ratio obtained from the pre-offline analysis of each fuel are vector-joined and aligned according to fuel type to obtain the dual-source calculation preparation feature set. Based on the inherent constant ratio between the atomic weight of carbon and the molecular weight of carbon dioxide, the molar mass ratio chemical equivalent conversion and decoupling identification assembly of the combination vectors of biomass sources and coal sources in the dual-source calculation preparation feature set are performed to obtain decoupled carbon emission data.

8. The carbon emission prediction method for biomass direct-hybrid power generation according to claim 1, characterized in that, Step S5 includes: Extract the historical evolution curves of fossil source carbon emissions from decoupled carbon emission data for several consecutive natural cycles over a preset period to construct a sliding observation window; The transient fossil carbon flow oscillation slope between adjacent time slices within the sliding observation window is calculated, and the original carbon release mass sequence and the oscillation slope are channel-stacked along the feature dimension to obtain the historical evolution tensor. A preliminary look-ahead prediction sequence is obtained by performing forward inference on the historical evolution tensor based on an attention-temporal convolution system. Based on the physical emission limit threshold determined by the exhaust oxygen concentration indicated by the tail probe in the aligned combustion state tensor and the total exhaust volume rate, the predicted values ​​of each future time slice node in the primary look-ahead prediction sequence are physically verified to obtain the carbon emission prediction sequence.

9. The carbon emission prediction method for biomass direct-hybrid power generation according to claim 4, characterized in that, The spatially segmented combustion tensor is obtained by asynchronously partitioning the spatial reaction zone of the ignition temperature field in the three-dimensional temperature cross-section data of the aligned combustion state tensor, including: Gradient adaptive continuous soft membership estimation is performed on the aligned combustion state tensor to obtain the adaptive soft membership tensor; Cross-coupling attribution correction is applied to the adaptive soft membership tensor to obtain the radiatively compensated membership tensor. We perform weighted soft segmentation based on the compensation membership tensor and the aligned combustion state tensor to obtain the spatially segmented combustion tensor.

10. A carbon emission prediction system for biomass direct hybrid power generation, characterized in that, include: The data spatiotemporal frequency alignment module is used to perform spatiotemporal frequency alignment processing on the acquired raw sensor data stream based on the flow field aerodynamic delay transmission module to obtain the aligned combustion state tensor; The dynamic feature extraction module is used to identify the reaction zone and extract asynchronous thermodynamic features from the aligned combustion state tensor to obtain the dynamic feature matrix. The carbon conversion rate decoupling inference module is used to perform asynchronous carbon conversion rate decoupling inference on the kinetic feature matrix based on the constraints of the embedded fuel ash conservation loss function to obtain the transient carbon conversion rate tuple; The stoichiometric decoupling conversion module is used to perform stoichiometric decoupling conversion on biomass sources and coal sources respectively, based on the transient carbon conversion rate tuple and the ratio of hysteresis compensation feed and carbon element content stored in the aligned combustion state tensor, to obtain decoupled carbon emission data. The trend evolution prediction module is used to perform trend evolution and forward prediction on decoupled carbon emission data through a temporal convolutional sequence network with an attention mechanism to obtain a carbon emission prediction sequence.

Images