A biomass gas coupled boiler nitrogen oxide prediction method and system

Abstract

The present application relates to the technical field of industrial prediction, and discloses a biomass gas coupled boiler nitrogen oxide prediction method and system, which first constructs a virtual transport coordinate system, generates coal powder and biomass gas energy packages, and performs trajectory tracking to obtain the time when the energy packages arrive at the furnace reaction zone, and constructs a feature reorganization matrix aligned with the reaction time; then a multi-model fusion prediction architecture containing a full-mixing flow reactor network mechanism model, a Gaussian process regression model and a physical information neural network is constructed, three types of prediction values are output based on the feature reorganization matrix, and the final NOx prediction value is obtained through weighted fusion; through energy package space-time tracking and feature reorganization, the present application effectively eliminates the feature misplacement problem caused by the differences in different fuel transport and reaction time lags, and improves the physical consistency and accuracy of nitrogen oxide prediction under variable load transient conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial forecasting technology, and more specifically, to a method and system for predicting nitrogen oxide emissions from a biomass gas coupled boiler. Background Technology

[0002] The coupled combustion technology of biomass gas and pulverized coal has become an important technological path for energy conservation and emission reduction in the thermal power industry due to its combination of the high efficiency of large-scale coal-fired power units and the low-carbon characteristics of biomass energy. In the coupled combustion process, nitrogen oxides (NOx), as one of the main pollutants, have a complex formation mechanism influenced by multiple factors such as fuel characteristics, combustion conditions, and the aerodynamic field within the furnace. To meet increasingly stringent environmental emission standards and achieve refined closed-loop control of the combustion process, establishing a high-precision, fast-response, and physically interpretable NOx emission prediction model has become a key technical challenge that urgently needs to be solved in this field.

[0003] In the prior art, some explorations have been made regarding NOx prediction for biomass boilers or coupled combustion systems. For example, Chinese patent application CN109359369A discloses a rapid calculation method for predicting NOx generation in a biomass stoker boiler. This method divides the grate into four physical zones: a preheating zone, a volatile matter release zone, a combustible carbon burnout zone, and a cooling zone. It also subdivides biomass fuel into components such as moisture, volatile matter, and combustible carbon. A gridded thermochemical balance calculation is used to simulate the gas chemical reactions within the furnace, thereby achieving rapid calculation of NOx generation in the stoker boiler. Furthermore, Chinese patent application CN120072119A discloses a high-efficiency, low-NOx optimization method and device for coupled biomass and coal combustion. This technical approach focuses on data-driven optimization. By acquiring the boiler's characteristic parameters, a genetic algorithm is used to optimize a BP neural network to construct a NOx emission prediction model. An efficiency model is then constructed using a support vector machine (SVM). Finally, a comprehensive cost calculation model is used to find the optimal set of operating parameters, addressing the problem of a single optimization object for boiler operating parameters.

[0004] However, while the aforementioned existing technologies have made some progress in parameter optimization under steady-state conditions or spatial partitioning calculations for specific furnace types, they still face the challenge of "spatiotemporal asynchrony" when dealing with the transient variable load process of coupled combustion of biomass gas and pulverized coal. Specifically, in coupled combustion systems, pulverized coal, as a solid-phase fluid, undergoes a series of long-cycle physicochemical processes, including pneumatic conveying through primary air ducts, in-furnace preheating, volatile matter release, and coke combustion; while biomass gas, as a gas-phase fluid, is conveyed through independent pipelines and undergoes rapid diffusion combustion after entering the furnace. The two have different transport medium flow rates, different pipeline physical lengths, and different time scales of combustion reaction mechanisms. Existing prediction methods, whether based on grid calculations of thermochemical equilibrium at a single moment or "snapshot" data-driven models that directly use the feeder speed and biomass gas valve opening collected by the DCS system at the same sampling time T as input features, all implicitly assume that "the data from each sensor are synchronized in real time." In reality, the "pulverized coal envelope" that participates in the furnace reaction to generate NOx at time T is emitted at time T-Δt1, while the "biomass gas envelope" that participates in the reaction at the same time is emitted at time T-Δt2. This simple "hard alignment of time" ignores the time lag differences in the physical transport and chemical reaction of heterogeneous fuels, leading to a misalignment in the causal relationship between input characteristics and output targets. Under transient conditions of rapid load increases or decreases or dynamic adjustments in fuel ratio, this misalignment will induce the model to learn incorrect input-output mapping relationships, ultimately resulting in phase lag in prediction results or distortion in dynamic response, failing to provide accurate feedforward guidance for real-time combustion optimization control. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, this invention provides a method and system for predicting nitrogen oxides (NOx) in biomass gas coupled boilers. It constructs a virtual transport coordinate system encompassing both pulverized coal and biomass gas channels. By generating energy envelopes and tracking their trajectories throughout their entire lifecycle, it accurately calculates the time delays of different fuels arriving at the furnace reaction zone, thereby constructing a physically and spatially aligned feature recombination matrix. Based on this, it integrates a fully mixed-flow mechanism model, Gaussian process regression, and a physical information neural network. This invention improves the physical consistency, robustness, and accuracy of NOx prediction under variable load transient conditions by eliminating feature misalignment caused by heterogeneous fuel transport and combining physical constraints with data correction.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting nitrogen oxide emissions from a biomass gas coupled boiler includes:

[0008] A virtual transport coordinate system is constructed, and energy packages are generated and tracked within the virtual transport coordinate system. The arrival time of the energy packages in the furnace reaction zone is obtained based on the trajectory tracking results. A feature recombination matrix aligned with the reaction time is constructed based on the arrival time. The energy packages include pulverized coal energy packages and biomass gas energy packages.

[0009] A multi-model fusion prediction architecture was constructed, comprising a fully mixed-flow reactor network mechanism model, a Gaussian process regression model, and a physical information neural network. Based on the feature recombination matrix, the theoretical NOx concentration was output using the fully mixed-flow reactor network mechanism model in the multi-model fusion prediction architecture. The theoretical NOx concentration was then corrected for deviation using the Gaussian process regression model to obtain the corrected NOx concentration. Finally, the physical information neural network was used for physical constraint prediction to obtain the predicted value.

[0010] The theoretical NOx concentration, the corrected NOx concentration, and the physical information neural network prediction value are weighted and fused to generate the final NOx prediction value.

[0011] The method for constructing the virtual transport coordinate system includes:

[0012] The physical parameters of the pulverized coal conveying path and the biomass gas conveying path are determined. Based on the physical parameters, the pulverized coal conveying path and the biomass gas conveying path are mapped to a virtual pulverized coal conveyor belt and a virtual biomass gas conveyor belt, respectively. A virtual transport coordinate system containing the virtual pulverized coal conveyor belt and the virtual biomass gas conveyor belt is established.

[0013] The origin of the pulverized coal virtual conveyor belt is set at the coal feeder outlet, and the endpoint is set at the furnace reaction zone inlet; the origin of the biomass gas virtual conveyor belt is set at the biomass gas valve, and the endpoint is set at the furnace reaction zone inlet.

[0014] The method for trajectory tracking includes:

[0015] The virtual conveyor belts for pulverized coal and biomass gas in the virtual transport coordinate system are spatially discretized to generate spatially discretized micro-elements.

[0016] Trajectory tracking of pulverized coal energy packages and biomass gas energy packages is performed in the spatial discretized infinitesimal elements of the virtual transport coordinate system, and the real-time position coordinates of the pulverized coal energy packages and biomass gas energy packages at each moment are calculated.

[0017] The method for determining whether the energy package has reached the furnace reaction zone includes:

[0018] Calculate the total length of the virtual conveyor belt path for pulverized coal and the virtual conveyor belt path for biomass gas; extract the position coordinates of the pulverized coal energy package and the position coordinates of the biomass gas energy package from the real-time position coordinates.

[0019] If the value corresponding to the position coordinate of the pulverized coal energy package is greater than or equal to the total length of the virtual pulverized coal conveyor path, it is determined that the pulverized coal energy package has reached the furnace reaction zone.

[0020] If the value corresponding to the position coordinates of the biomass gas energy package is greater than or equal to the total length of the virtual biomass gas conveyor belt path, then it is determined that the biomass gas energy package has reached the furnace reaction zone.

[0021] The method for constructing the feature recombination matrix includes:

[0022] Based on the current reaction time in the furnace, among all energy packages that have completed trajectory tracking, search for pulverized coal energy packages and biomass gas energy packages whose arrival time matches the current reaction time; extract the component label vector attributes and pulverized coal mass flow rate of the matching pulverized coal energy package, as well as the component label vector attributes and biomass gas volume flow rate of the matching biomass gas energy package, and combine them with the furnace operating parameters at the current reaction time to construct a feature recombination matrix aligned with the reaction time.

[0023] The method for constructing the network mechanism model of the fully mixed-flow reactor includes:

[0024] The furnace space is divided into three series-connected fully mixed-flow reactors: the main combustion zone, the reduction zone, and the burnout zone. The component label vector attributes of pulverized coal energy and the pulverized coal mass flow rate, as well as the component label vector attributes of biomass gas energy and the biomass gas volume flow rate in the feature recombination matrix, are used as the inlet boundary conditions of the fully mixed-flow reactor in the main combustion zone to construct a network mechanism model of the fully mixed-flow reactor.

[0025] The input vector of the Gaussian process regression model is composed of a feature recombination matrix and the theoretical NOx concentration; the Gaussian process regression model outputs the mean prediction deviation and the prediction variance, and the mean prediction deviation is added to the theoretical NOx concentration to obtain the corrected NOx concentration.

[0026] The physical information neural network is trained by minimizing the total loss function, which is a weighted sum of the defined physical constraint loss and data loss.

[0027] The physical constraint loss is composed of the sum of the mass conservation constraint loss, the energy conservation constraint loss, and the stoichiometric constraint loss.

[0028] A nitrogen oxide prediction system for a biomass gas coupled boiler, used to implement the above-mentioned nitrogen oxide prediction method for a biomass gas coupled boiler, the system comprising:

[0029] Feature Recombination Module: Used to construct a virtual transport coordinate system, generate energy packages in the virtual transport coordinate system and perform trajectory tracking, obtain the arrival time of the energy packages to the furnace reaction zone based on the trajectory tracking results, and construct a feature recombination matrix aligned with the reaction time based on the arrival time; the energy packages include pulverized coal energy packages and biomass gas energy packages;

[0030] The multi-model prediction module is used to construct a multi-model fusion prediction architecture that includes a fully mixed-flow reactor network mechanism model, a Gaussian process regression model, and a physical information neural network. Based on the feature recombination matrix, it uses the fully mixed-flow reactor network mechanism model in the multi-model fusion prediction architecture to output the theoretical NOx concentration, uses the Gaussian process regression model in the multi-model fusion prediction architecture to correct the deviation of the theoretical NOx concentration, and uses the physical information neural network in the multi-model fusion prediction architecture to perform physical constraint prediction, obtaining the predicted value of the physical information neural network.

[0031] Weighted fusion module: Used to weight and fuse the theoretical NOx concentration, the corrected NOx concentration, and the physical information neural network prediction value to generate the final NOx prediction value.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] This invention constructs a virtual transport coordinate system and tracks the trajectory of pulverized coal and biomass gas energy envelopes, accurately quantifying the time delay differences of different forms of fuel along the physical transport path. The feature recombination matrix constructed based on arrival time effectively eliminates the feature misalignment problem caused by different medium flow rates and reaction characteristics in traditional time alignment methods, ensuring strict spatiotemporal synchronization of input features in terms of physicochemical reaction causality. Combining the establishment of physical benchmarks for the fully mixed-flow reactor network mechanism model, the nonlinear deviation correction of the Gaussian process regression model, and the conservation law constraints of the physical information neural network, this multi-model fusion architecture not only uses the mechanism model to ensure the trend correctness of the prediction results, but also overcomes the defects of single models in terms of weak generalization ability or violation of physical laws under variable load transient conditions through a dual correction mechanism of data-driven and physical constraints. Thus, it achieves high-precision, low-latency prediction of nitrogen oxide emissions from biomass gas coupled boilers while ensuring physical consistency. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 A flowchart of a method for predicting nitrogen oxides in a biomass gas coupled boiler, provided by an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the virtual transport coordinate system provided in an embodiment of the present invention;

[0037] Figure 3 A schematic diagram illustrating the timing differences between pulverized coal and biomass gas during transmission, provided for an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of a multi-model fusion prediction architecture provided in an embodiment of the present invention;

[0039] Figure 5 This is a functional block diagram of a biomass gas coupled boiler nitrogen oxide prediction system provided in an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1:

[0042] Please see Figure 1 As shown, this embodiment provides a method for predicting nitrogen oxide emissions from a biomass gas coupled boiler, including:

[0043] Step S10: Construct a virtual transport coordinate system including a virtual conveyor belt for pulverized coal and a virtual conveyor belt for biomass gas. Generate energy packages in the virtual transport coordinate system and perform trajectory tracking. Obtain the arrival time of the energy packages to the furnace reaction zone based on the trajectory tracking results. Construct a feature recombination matrix aligned with the reaction time based on the arrival time. The energy packages include pulverized coal energy packages and biomass gas energy packages.

[0044] Specifically, in a biomass gas coupled coal-fired boiler system, pulverized coal, as a solid particulate fuel, is transported to the furnace through a primary air duct, while biomass gas, as a gaseous fuel, is transported to the furnace through a separate gas pipeline. The difference in the physical forms of the two fuels determines the fundamental difference in their flow characteristics in their respective transport pipelines. When pulverized coal particles move under the influence of primary air, due to the interaction between particle inertia and airflow resistance, the actual velocity of the pulverized coal particles is lower than the velocity of the carrier gas. This velocity difference is called the slip effect in fluid mechanics. Biomass gas, as a continuous phase gas, has a flow velocity directly related to the pressure gradient and flow cross-sectional area within the pipeline, and there is no slip effect during the flow process. In addition to the difference in physical transport velocity, the chemical reaction characteristics of the two fuels after entering the furnace also differ significantly: the combustible components in biomass gas, including hydrogen, carbon monoxide, and methane, can complete the combustion reaction within milliseconds after contact with oxygen, while pulverized coal particles need to undergo multiple series processes such as preheating, volatile matter release, volatile matter combustion, and coke heterogeneous combustion, with the entire combustion process taking place on the order of seconds. The combined effect of the aforementioned differences in transmission speed and reaction time delay leads to a situation where the coal powder that actually participates in combustion and generates nitrogen oxides at a certain moment in the furnace reaction zone leaves the coal feeder at a different time than the biomass gas that participates in the reaction at the same moment leaves the valve. If traditional data processing methods are used, directly using the coal feeder speed signal and biomass gas valve opening signal acquired by the distributed control system (DCS) at the same sampling moment as the input features of the prediction model, then the coal powder flow characteristics and biomass gas flow characteristics input to the model do not correspond to the same reaction event in physical space and time. The model will learn based on misaligned causal relationships, and the prediction results will show obvious time phase deviations, which are particularly significant under transient conditions such as rapid load changes.

[0045] Step S10 abstracts the pulverized coal conveying path and the biomass gas conveying path into two independent virtual conveyor belts, establishes a unified virtual transport coordinate system, and discretizes the continuously flowing fuel into energy packages carrying complete physicochemical properties. Using fluid dynamics principles, the trajectory of each energy package on the virtual conveyor belt is tracked in real time, and the transport delay time from entering the system to reaching the furnace reaction zone is accurately calculated. Based on this, using the current reaction time in the furnace as a benchmark, the fuel energy packages that actually participate in the reaction at that time are searched back, their composition and flow information are extracted, and combined with the furnace operating parameters to construct a feature recombination matrix that is strictly aligned in physical space and time. This processing method transforms the complex three-dimensional two-phase flow problem into a computable one-dimensional transport tracking problem, enabling the input features of the prediction model to have a real causal correspondence at the physical level, providing a physically self-consistent data foundation for the multi-model fusion prediction in the subsequent step S20.

[0046] Further, step S10 includes,

[0047] Step S11: Determine the physical parameters of the pulverized coal conveying path and the biomass gas conveying path. Based on the physical parameters, map the pulverized coal conveying path and the biomass gas conveying path to a virtual pulverized coal conveyor belt and a virtual biomass gas conveyor belt, respectively, and establish a virtual transport coordinate system including the virtual pulverized coal conveyor belt and the virtual biomass gas conveyor belt. The origin of the coordinate system of the virtual pulverized coal conveyor belt is set at the coal feeder outlet, and the endpoint is set at the furnace reaction zone inlet. The origin of the coordinate system of the virtual biomass gas conveyor belt is set at the biomass gas valve, and the endpoint is set at the furnace reaction zone inlet.

[0048] Please see Figure 2 This diagram schematically illustrates the virtual transport coordinate system constructed in this embodiment, visually presenting the parallel architecture and spatial mapping relationship of the pulverized coal virtual conveyor belt and the biomass gas virtual conveyor belt. Specifically, the physical parameters of the pulverized coal conveying path include the actual length of the primary air duct from the coal feeder outlet to the furnace burner and the length of the furnace preheating section before the pulverized coal reaches the ignition temperature after entering the furnace. The physical parameters of the biomass gas conveying path include the actual length of the gas duct from the biomass gas regulating valve to the furnace nozzle and the length of the virtual diffusion section for rapid diffusion and mixing after the gas enters the furnace. The total path length Lcoal of the pulverized coal virtual conveyor belt is equal to the sum of the actual length of the primary air duct and the length of the preheating section in the furnace. The preheating section length is introduced to characterize the spatial distance corresponding to the preheating process of pulverized coal particles from entering the furnace to the start of a vigorous chemical reaction. This length depends on the particle size distribution characteristics of the pulverized coal and the primary air temperature. Smaller particle sizes result in larger specific surface areas and faster heating rates; higher primary air temperatures result in higher initial particle temperatures and less heat absorption required to reach the ignition point. Both factors jointly determine the value of the preheating section length. The total path length Lgas of the biomass gas virtual conveyor belt is equal to the sum of the actual length of the gas duct and the length of the virtual diffusion section. The virtual diffusion section length is introduced to characterize the equivalent spatial distance corresponding to the rapid mixing process of gaseous fuel after entering the furnace due to the much higher molecular diffusion rate than solid particles. This length mainly depends on the gas nozzle structure and the turbulence intensity in the furnace. For example, for a conventional nozzle structure, the virtual diffusion section length is approximately 0.8 meters.

[0049] Both the pulverized coal virtual conveyor belt and the biomass gas virtual conveyor belt are defined as a one-dimensional spatial coordinate system. Figure 2 The horizontal, one-way arrow pointing to the right indicates the positive direction of the one-dimensional spatial coordinate x. The origin of the coordinate system is set at the starting position of fuel entering the system, i.e., the coal feeder outlet or the biomass gas regulating valve, and the endpoint is set at the inlet of the furnace reaction zone. Figure 2 The common endpoints marked in the coordinate system are aligned with the positive direction of the fuel flow, pointing towards the furnace. This method of defining a one-dimensional coordinate system combines... Figure 2The linearized representation simplifies the complex pipeline path in actual three-dimensional space into a single variable of distance along the pipeline. This allows subsequent position tracking calculations to focus only on the displacement of fuel along the flow direction, avoiding the enormous computational resource consumption required for three-dimensional flow field calculations. Two virtual conveyor belts together constitute a virtual transport coordinate system, which is based on... Figure 2 The furnace reaction zone inlet shown is a common spatial endpoint, enabling time synchronization determination of fuels transported via two different paths within a unified spatial reference frame. Step S11 maps the respective physical transport paths of pulverized coal and biomass gas as follows: Figure 2 The system uses a one-dimensional virtual conveyor belt with clearly defined start and end points and length parameters. This mapping process allows the fuel position in continuous space to be quantitatively expressed using a single coordinate value, providing a geometric basis for the spatial discretization process in step S12 and a spatial reference for the trajectory tracking calculation in step S14. Without the virtual transport coordinate system established in step S11, subsequent steps would be unable to quantitatively describe the fuel position, the trajectory tracking calculation would lose its spatial reference frame, the displacement integral result of the energy envelope would not be able to be compared with the arrival judgment condition, and the entire spatiotemporal tracking mechanism would not function.

[0050] Step S12: Spatial discretization processing is performed on the virtual conveyor belt for pulverized coal and the virtual conveyor belt for biomass gas in the virtual transport coordinate system to generate spatially discretized micro-elements.

[0051] Specifically, spatial discretization refers to dividing a continuous one-dimensional virtual conveyor belt along its length into several interconnected micro-segments. Each micro-segment has a fixed spatial length, and the segment numbers increase sequentially in the positive direction starting from the origin. The purpose of discretization is to transform the position tracking problem in continuous space into a problem of changing micro-segment numbers in discrete space, facilitating the discretized recording and updating of the energy envelope's position during numerical calculations. The granularity of the spatial discretization micro-elements is dynamically adjusted according to the system's current operating conditions: when the system is in transient conditions, the airflow and feed rate change significantly in a short time, and the fuel flow rate fluctuates drastically. In this case, a finer discretization granularity is needed to accurately capture the impact of flow rate changes on position; when the system is in steady-state conditions, the airflow and feed rate remain relatively constant, and the fuel flow rate changes gradually. In this case, a coarser discretization granularity can be used to reduce computational resource consumption.

[0052] The determination of transient and steady-state operating conditions is based on the load change rate index: the current electrical load value of the boiler and the electrical load value of the previous minute are obtained, and the absolute value of the difference between the two is calculated. If the difference exceeds the load change rate threshold Pthreshold, the system is determined to be in a transient operating condition; otherwise, it is determined to be in a steady-state operating condition. For example, the load change rate threshold Pthreshold can be set to five megawatts per minute. The basis for setting this threshold is: when the load change rate exceeds this value, the adjustment range of the primary air volume and biomass gas flow will cause observable changes in the fuel flow rate within a single control cycle, requiring improved spatial tracking accuracy; when the load change rate is lower than this value, the impact of flow rate changes on tracking accuracy is within the engineering acceptable range, and the discretization granularity can be appropriately relaxed. When the system is in transient operation, the virtual pulverized coal conveyor belt is divided into Nfine micro-segments, each with a length of Lcoal / Nfine. Similarly, the virtual biomass gas conveyor belt is divided into Nfine micro-segments, each with a length of Lgas / Nfine. For example, Nfine can be taken as one hundred. When the system is in steady-state operation, both virtual conveyor belts are divided into Ncoarse micro-segments each. For example, Ncoarse can be taken as twenty. The segment lengths increase accordingly, and the computational complexity is reduced to one-fifth of that in the transient operation.

[0053] Step S12's spatial discretization transforms position tracking in continuous space into updating the numbers within discrete micro-elements, optimizing computational efficiency while maintaining tracking accuracy. The strategy of dynamically adjusting the discretization granularity allows the system to maintain high tracking accuracy under transient conditions and reduce computational burden under steady-state conditions, achieving an adaptive balance between accuracy and efficiency. The spatially discretized micro-elements generated in step S12 provide a discretized computational framework for trajectory tracking in step S14: the displacement increment encapsulated within each integration step is compared with the micro-element length to determine whether the boundary of the micro-element has been crossed. When the cumulative displacement exceeds the total path length, an arrival determination is triggered.

[0054] Step S13: Discretize the continuously flowing pulverized coal and biomass gas into pulverized coal energy packages and biomass gas energy packages according to time intervals, and assign timestamp attributes and component label vector attributes to the pulverized coal energy packages and biomass gas energy packages; the timestamp attribute records the generation time when the energy package enters the origin of the virtual conveyor belt coordinates.

[0055] Specifically, an energy package is a discretized computational unit defined for achieving spatiotemporal tracking of fuel. Each energy package represents a clump of fuel that passes through the system inlet within a specific time interval. This fuel clump is considered an independent tracking object carrying time information, mass or volume information, and chemical composition information. Discretizing the continuous flow of fuel into energy packages allows the system to track a specific clump of fuel entering the system at a specific time, recording its complete spatiotemporal trajectory from entry into the system to arrival at the reaction zone. This is the foundation for achieving reaction time alignment.

[0056] The process of generating pulverized coal energy packages is as follows: Using a time interval Δtcoal as the period, at the end of each period, the mass of pulverized coal passing through the feeder outlet within that period is packaged into an independent pulverized coal energy package. The time interval Δtcoal is dynamically adjusted according to the feeder speed. The adjustment logic is as follows: the higher the feeder speed, the greater the mass of pulverized coal output per unit time. To maintain the mass of pulverized coal represented by each energy package within a reasonable range, the generation interval needs to be shortened; the lower the feeder speed, the longer the generation interval can be appropriately extended. For example, when the feeder speed is n_coal revolutions per minute, the time interval Δtcoal can be set as the ratio of the base interval Δtbase_coal multiplied by the reference speed n_ref divided by the current speed n_coal. For a working condition with a speed of 30 revolutions per minute, if the base interval is 2 seconds and the reference speed is 30 revolutions per minute, then Δtcoal equals 2 seconds. The generation process of biomass gas energy envelopes is similar: Using a time interval Δtgas as the period, at the end of each period, the gas volume passing through the biomass gas regulating valve during that period is packaged into an independent biomass gas energy envelope. The time interval Δtgas is dynamically adjusted according to the valve opening change rate. The adjustment logic is as follows: a larger valve opening change rate indicates that the gas flow is being rapidly regulated, requiring a shorter generation interval to accurately capture flow changes; a smaller valve opening change rate indicates relatively stable flow, allowing for a more extended generation interval. For example, when the valve opening change rate is less than or equal to 5% per second, Δtgas is set to one second; when the valve opening change rate exceeds 5% per second, Δtgas is shortened to 0.2 seconds.

[0057] Each generated energy package is assigned a timestamp attribute Tbirth, which records the precise moment the energy package enters the virtual conveyor belt's coordinate origin, i.e., the coal feeder outlet or biomass gas regulating valve. The time accuracy should be consistent with or higher than the DCS system's sampling period. The timestamp attribute is essential information for subsequent calculations of transmission delay: by comparing the energy package's generation time Tbirth with its arrival time Tarrival, the package's dwell time on the transport path can be obtained.

[0058] Each energy package is assigned a component label vector attribute, which records the chemical composition information of the fuel represented by the energy package in vector form. The component label vector Ccoal for pulverized coal energy packages contains five components: volatile matter content, fixed carbon content, moisture content, ash content, and nitrogen content. These component parameters directly determine the combustion characteristics of pulverized coal and the potential for the formation of fuel-type nitrogen oxides. Volatile matter content affects the ignition rate and flame stability of pulverized coal; fixed carbon content affects the reaction duration of the coke combustion stage; moisture content affects the furnace temperature distribution and the heat load of the reaction zone; ash content affects furnace ash accumulation and heat exchange efficiency of the heating surface; and nitrogen content directly determines the total amount of nitrogen in the fuel that can be converted into nitrogen oxides. The component data comes from the real-time measurement results of an online coal quality analyzer. If the online coal quality analyzer is unavailable, typical values ​​matched with the current coal type in the historical coal quality database are used. The component label vector Cgas for biomass gas energy packages contains five components: hydrogen concentration, carbon monoxide concentration, methane concentration, carbon dioxide concentration, and gas calorific value. Hydrogen and carbon monoxide, as reducing gases, can react with nitrogen oxides in the reduction zone of the furnace to generate nitrogen, and their concentrations directly affect the nitrogen oxide reduction efficiency. Methane concentration affects the flame temperature of the biomass gas. Carbon dioxide, as an inert component, dilutes the concentration of combustible gases. The calorific value of the gases determines the energy contribution ratio of the biomass gas. The composition data comes from real-time measurements by a gas composition analyzer at the gasifier outlet.

[0059] Step S13 discretizes the continuous fuel flow into traceable, independent energy packets, assigning complete temporal and compositional information to each packet. This allows subsequent trajectory tracking to not only determine the spatial location changes of the fuel but also preserve its chemical properties for use in prediction models. The assignment of a timestamp attribute gives each energy packet a unique time identifier, forming the information basis for calculating the transmission delay in step S15. The assignment of a component tag vector attribute enables step S16 to extract the component information of the reacting fuels and construct a feature recombination matrix. The strategy of dynamically adjusting the energy packet generation interval allows the system to control the number of energy packets while ensuring tracking accuracy, avoiding wasted computational resources due to an excessive number of packets and insufficient temporal resolution due to an insufficient number of packets.

[0060] Step S14: Track the pulverized coal energy package and the biomass gas energy package in the spatial discretized micro-element of the virtual transport coordinate system, and calculate the real-time position coordinates of the pulverized coal energy package and the biomass gas energy package at each time.

[0061] Specifically, trajectory tracking refers to the real-time calculation of the motion state of each energy envelope within a virtual conveyor belt coordinate system, based on fluid dynamics principles, to determine its position coordinates at any given time. The core of trajectory tracking lies in establishing the motion equations of the energy envelopes and solving for the relationship between position and time through numerical integration.

[0062] The equation of motion for the coal powder energy envelope is established as follows: Coal powder particles flow along the pipe under the carrying of primary air. Due to the relative motion between the particles and the airflow, the particles are accelerated by the drag force exerted by the airflow. At the same time, the inertia of the particles causes their velocity to lag behind the airflow velocity; the velocity difference between the two is called slip. Under steady-state flow conditions, the particle acceleration approaches zero, and a stable proportional relationship is formed between the particle velocity and the airflow velocity; this ratio is called the slip coefficient. The instantaneous velocity Vcoal(t) of the coal powder energy envelope is equal to the primary air velocity Vprimary(t) multiplied by the slip coefficient αslip, i.e., Vcoal(t) = αslip × Vprimary(t). The primary air velocity Vprimary(t) is calculated by dividing the measured primary air volume by the cross-sectional area of ​​the primary air pipe. The slip coefficient αslip is determined by Stokes' law from the average particle size and density of the coal powder particles: the larger the particle size, the greater the particle inertia, the more significant the slip effect, and the smaller the slip coefficient; the smaller the particle size, the better the particle flowability, and the closer the slip coefficient is to one. For example, for coal powder particles with an average particle size of 75 micrometers, the slip coefficient αslip is approximately 0.85.

[0063] The equation of motion for the biomass gas energy envelope is established as follows: As a continuous phase gas flowing in the pipeline, biomass gas exhibits no slip effect, and the instantaneous velocity Vgas(t) of the energy envelope is equal to the average velocity within the pipeline. The average velocity within the pipeline is calculated by dividing the volumetric flow rate Qgas(t) measured by the mass flow meter by the cross-sectional area Agas of the gas pipeline.

[0064] Based on the aforementioned equations of motion, a numerical integration method is used to calculate the real-time position coordinates of the energy envelope. The position coordinate Pos(t) is calculated by integrating the instantaneous velocity over time: in a discrete-time frame, the time axis is divided into several integration steps Δtint. Within each integration step, the displacement increment ΔPos is calculated based on the instantaneous velocity, and the accumulated displacement increments yield the real-time position coordinates. For example, the integration step Δtint can be set to 0.1 seconds. Since the instantaneous velocity of the energy envelope is calculated from the airflow or valve opening signals collected in real time by the DCS system, and velocity is an explicit function of time rather than position, the position integration problem is simplified to a numerical integration problem of a known time-varying velocity function. The integration method used is the fourth-order Runge-Kutta method. This method calculates multiple intermediate slope values ​​within each integration step and then performs a weighted average to achieve a high-precision estimate of the displacement increment, exhibiting higher numerical stability and accuracy compared to the simple Euler method. The calculation process of the fourth-order Runge-Kutta method is as follows: Let the current time be ti and the position coordinate be Pos(ti). First, calculate four intermediate velocity values ​​k1, k2, k3, and k4, where k1 is the instantaneous velocity at time ti and position Pos(ti), k2 is the velocity at time ti + Δtint / 2 and position Pos(ti) + k1 × (Δtint / 2), k3 is the velocity at time ti + Δtint / 2 and position Pos(ti) + k2 × (Δtint / 2), and k4 is the velocity at time ti + Δtint and position Pos(ti) + k3 × Δtint. For the case where k2 and k3 are the same velocity at the same time, the Runge-Kutta method improves the integration accuracy through different weighting methods. The displacement increment ΔPos = Δtint × ((k1 + 2 × k2 + 2 × k3 + k4) / 6); the position coordinate at the next time step Pos(t(i+1)) = Pos(ti) + ΔPos. When the sampling period of the DCS system is greater than the integration step size, the velocity value at the intermediate time is obtained by linearly interpolating the velocities at adjacent sampling times.

[0065] For example, set the current time. The real-time position coordinates of the pulverized coal energy wrapped on the virtual pulverized coal conveyor belt. Meters, the system-defined integration step size The slip coefficient is calculated based on the average particle size of the pulverized coal. Assume the primary wind speed values ​​obtained by the DCS system and after linear interpolation within the integration interval are as follows: The primary wind speed at 20.00 m / s is 20.00 m / s. The primary wind speed was 20.20 m / s. The primary wind speed at 20.40 m / s is given. The calculation process first determines four intermediate velocity values: k1 is... The instantaneous velocity of the coal powder energy envelope at any given moment, i.e. k2 and k3 correspond to the midpoint of the integral. The speed in seconds, that is k4 is the endpoint time. The speed in seconds, that is Then, the displacement increment within this integration step is calculated according to the fourth-order Runge-Kutta weighting formula. Substituting the numerical values, we get Finally, update the next time step t(i+1). Real-time position coordinates in seconds The calculation process iterates within each time step until the position coordinate value is greater than the total length of the virtual conveyor belt path for pulverized coal.

[0066] Step S14 establishes motion equations based on fluid dynamics principles and solves them using a high-precision numerical integration method, achieving precise tracking of the energy package's trajectory. The introduction of the slip coefficient ensures that the velocity calculation of the pulverized coal energy package reflects the physical interaction between the particles and the carrier gas, unlike the simplistic assumption of directly equating particle velocity with airflow velocity, thus improving the accuracy of transmission time estimation. The use of the fourth-order Runge-Kutta method significantly reduces accumulated errors compared to lower-order integration methods, ensuring that the energy package's position coordinates remain reliably accurate even after long-term tracking. The trajectory tracking results of step S14 provide positional data for the arrival determination and transmission delay time calculation in step S15: only by obtaining the precise position coordinates of the energy package at each moment can it be determined whether it has reached the furnace reaction zone and the arrival time recorded. Without the trajectory tracking in step S14, the energy package's position cannot be determined, step S15 cannot perform the arrival determination, the transmission delay time cannot be calculated, and the entire spatiotemporal alignment mechanism will fail.

[0067] Step S15: Determine whether the pulverized coal energy package and the biomass gas energy package have reached the furnace reaction zone based on the real-time location coordinates. If they have reached the furnace reaction zone, record the arrival time. Calculate the transmission delay time of the pulverized coal energy package and the transmission delay time of the biomass gas energy package based on the arrival time and the generation time.

[0068] Specifically, the arrival determination logic is as follows: after each trajectory tracking calculation cycle, the current position coordinates Pos(t) of the energy package are compared with the total length of the corresponding virtual conveyor path. For pulverized coal energy packages, if the value corresponding to its position coordinates is greater than or equal to the total length Lcoal of the pulverized coal virtual conveyor path, then the pulverized coal energy package is determined to have reached the furnace reaction zone; for biomass gas energy packages, if the value corresponding to its position coordinates is greater than or equal to the total length Lgas of the biomass gas virtual conveyor path, then the biomass gas energy package is determined to have reached the furnace reaction zone. The reason for directly comparing the coordinate values ​​with the total path length is that both the pulverized coal virtual conveyor and the biomass gas virtual conveyor are defined as a one-dimensional spatial coordinate system. This one-dimensional coordinate system definition simplifies the complex pipeline routing in actual three-dimensional space into a single variable of distance along the path, so that the coordinate value of any point on the virtual conveyor represents the cumulative transport distance of that point relative to the starting position.

[0069] When an energy package is determined to have reached the furnace reaction zone, this moment is recorded as the arrival time, Tarrival. The time precision of the arrival time is consistent with the precision of the integration step size in trajectory tracking. Since the actual position update is discrete, the precise time when the energy package crosses the path endpoint may fall within the middle of a certain integration step size. In this case, the precise arrival time can be estimated using a linear interpolation method: Assume that at time ti, the position coordinate Pos(ti) is less than the total path length L, and at time t(i+1), the position coordinate Pos(t(i+1)) is greater than or equal to L. Then the precise arrival time... .

[0070] The formula for calculating the transmission delay time (Delay) is: Delay = Tarrival - Tbirth. The transmission delay time characterizes the time interval from when fuel enters the conveying system, including the coal feeder outlet or biomass gas valve, and reaches the furnace reaction zone. This time interval is affected by factors such as the conveying path length, fuel flow rate, and flow rate variation history. Because pulverized coal and biomass gas have different conveying path lengths and flow rate characteristics, even if pulverized coal energy envelopes and biomass gas energy envelopes are generated at the same time, their arrival times in the furnace reaction zone will differ, resulting in a difference in their transmission delay times. For example, see [link to example]. Figure 3 , Figure 3 The time-series differences between pulverized coal and biomass gas during transmission are visually displayed using a timeline. For a pulverized coal energy package generated at a certain generation time T0, the transmission delay time Delay is shown. coal The time delay is likely 15 seconds, indicating that the package arrives at the furnace reaction zone at time T0+15 seconds, corresponding to the time marked "arrival after pulverized coal" in the diagram; the biomass gas energy package generated at the same time T0 has a transmission delay time Delay. gasThe time difference might be 3 seconds, indicating that the package arrived at the furnace reaction zone at time T0+3 seconds, corresponding to the "biomass gas arrives first" time marked in the diagram. There is a 12-second delay between the two. If the traditional method of aligning data at the same time is used, employing the pulverized coal flow rate and biomass gas flow rate at time T0 as input features to predict the nitrogen oxide concentration at time T0, then the pulverized coal feature input to the model actually corresponds to the pulverized coal entering the system at time T0-15 seconds, and the biomass gas feature actually corresponds to the gas entering the system at time T0-3 seconds. These two are not participants in the same reaction event in physical space and time, resulting in a misalignment in the causal relationship learned by the model.

[0071] Step S16: Based on the current reaction time in the furnace, search for pulverized coal energy packages and biomass gas energy packages whose arrival time matches the current reaction time among all energy packages that have completed trajectory tracking; extract the component label vector attributes and pulverized coal mass flow rate of the matching pulverized coal energy package, as well as the component label vector attributes and biomass gas volume flow rate of the matching biomass gas energy package, and construct a feature recombination matrix aligned with the reaction time by combining the furnace operating parameters at the current reaction time.

[0072] Specifically, the current reaction time (Treaction) refers to the moment when combustion and nitrogen oxide (NOx) generation are currently occurring in the furnace reaction zone, which is the target moment for NOx concentration prediction. The purpose of backtracking based on the reaction is to determine when the fuel truly participating in the furnace reaction originates from the energy package entering the delivery system, thereby extracting the correct fuel characteristics as input to the prediction model.

[0073] The matching search process is as follows: Iterate through all energy packages with completed trajectory tracking, including pulverized coal energy packages and biomass gas energy packages. For each energy package, check whether its arrival time (Tarrival) matches the current reaction time (Treaction). A time window tolerance mechanism is used for matching: if the absolute value of the difference between Tarrival and Treaction is less than or equal to the matching tolerance threshold εmatch, then the energy package is considered to match the current reaction time. The matching tolerance threshold εmatch is determined based on the energy package generation interval and the accuracy of the integration step. For example, it can be set to half of the energy package generation interval to ensure that at least one package can match the target reaction time within adjacent generation intervals. Packages with matching arrival times in the pulverized coal energy package set are denoted as matched pulverized coal energy packages; packages with matching arrival times in the biomass gas energy package set are denoted as matched biomass gas energy packages.

[0074] When a certain type of fuel has no matching energy package at the target reaction time, for example due to a time gap caused by the energy package generation interval, a linear interpolation method is used to estimate the virtual energy package information at that time. The interpolation process is as follows: In the set of energy packages for that type of fuel, search for the two energy packages whose arrival time is closest to the current reaction time (Treaction) and which are located before and after the current reaction time (Treaction), respectively, and denoted as the front package and the back package; using the arrival time as the interpolation variable, perform linear interpolation on the component label vectors and flows of the front and back packages to calculate the virtual components and virtual flows at the current reaction time (Treaction).

[0075] Information from matching energy packages is extracted and combined with furnace operating parameters to construct a feature recombination matrix, Mreaction. The structure of the Mreaction matrix is ​​designed as follows: the first row vector contains the components of the component label vector Ccoal for matching pulverized coal energy packages and the corresponding pulverized coal mass flow rate mcoal; the second row vector contains the components of the component label vector Cgas for matching biomass gas energy packages and the corresponding biomass gas volumetric flow rate vgas; the third row vector contains the furnace operating parameters corresponding to the current reaction time Treaction, specifically including secondary air volume, burnout air volume, furnace temperature, and furnace oxygen content. These furnace operating parameters are directly obtained from the real-time measurements of Treaction at the current reaction time from the DCS system. Min-max normalization or standard fraction normalization is used to normalize all elements in the Mreaction matrix to eliminate dimensional differences.

[0076] The feature recombination matrix Mreaction constructed in step S16 achieves strict physical-temporal alignment of the input features: the coal powder composition and flow rate information in the matrix comes from the coal powder energy package arriving at the furnace reaction zone at the Treaction time, the generation time of which is Treaction minus the coal powder transport delay time; the biomass gas composition and flow rate information in the matrix comes from the biomass gas energy package arriving at the furnace reaction zone at the Treaction time, the generation time of which is Treaction minus the biomass gas transport delay time; the furnace operating parameters in the matrix are real-time measurements at the Treaction time. These three types of information physically correspond to the same reaction event, namely the combustion process occurring in the furnace reaction zone at the Treaction time, the amount of nitrogen oxides generated in this process being jointly determined by the fuel composition participating in the reaction and the furnace environmental conditions. This spatiotemporally aligned feature combination method differs from the traditional simple stacking of data from the same sampling time, enabling the input features of the prediction model to possess a true physical causal relationship.

[0077] Step S10 transforms the raw sensor signals collected by the distributed control system into a feature reconstruction matrix reflecting physical reaction events. This data preprocessing is completed before the multi-model fusion prediction in step S20 to ensure that the features entering the prediction model have physical consistency in the spatiotemporal dimension. Without the spatiotemporal tracking and feature reconstruction processing in step S10, the prediction model in step S20 would directly use sensor data from the same sampling time as input. Due to the different transmission delays of pulverized coal and biomass gas, the input features correspond to fuels entering the system at different times, resulting in a spatiotemporal misalignment with the actual reaction state in the furnace reaction zone. The model training process would learn incorrect causal relationships, associating fuel features that do not participate in the reaction simultaneously with the amount of nitrogen oxides generated, leading to phase deviations in the prediction results under transient conditions. By introducing a virtual conveyor belt model and an energy package tracking mechanism, not only is the spatiotemporal misalignment problem of traditional feature alignment methods corrected, but a complete spatiotemporal archive of the fuel transportation process is also provided to the system. This archive records the generation time, composition properties, trajectory, and arrival time of each energy package, enabling the system to perform retrospective analysis of the fuel transportation process. It can trace the source of fuel with abnormal nitrogen oxide concentrations at specific times, providing data support for boiler combustion diagnostics and raw material quality control. Real-time calculation of transmission delay time can also be used to indirectly monitor the operating status of the fuel delivery system. When the calculated transmission delay time deviates continuously from historical benchmark values, it may indicate equipment abnormalities such as pipe ash accumulation, decreased fan efficiency, or sluggish valve response, providing a reference indicator for preventative maintenance.

[0078] Step S20, see Figure 4 A multi-model fusion prediction architecture was constructed, comprising a fully mixed-flow reactor network mechanism model, a Gaussian process regression model, and a physical information neural network. Based on the feature recombination matrix, the theoretical NOx concentration was output using the fully mixed-flow reactor network mechanism model in the multi-model fusion prediction architecture. The theoretical NOx concentration was then corrected for deviation using the Gaussian process regression model in the multi-model fusion prediction architecture to obtain the corrected NOx concentration. Physical constraint prediction was performed using the physical information neural network in the multi-model fusion prediction architecture to obtain the predicted value of the physical information neural network. Finally, the theoretical NOx concentration, the corrected NOx concentration, and the predicted value of the physical information neural network were weighted and fused to generate the final NOx prediction value.

[0079] Specifically, the formation of nitrogen oxides (NOx) in biomass gas coupled with coal-fired boilers involves a complex thermochemical reaction network. The formation pathways of NOx include two main mechanisms: thermal NOx and fuel-based NOx. Thermal NOx is formed by the reaction of nitrogen in the air with oxygen under high-temperature conditions, and its formation rate is exponentially related to the furnace temperature. Fuel-based NOx is formed by the oxidation of organic nitrogen in the fuel during combustion, and its formation amount is closely related to the nitrogen content of the fuel and the combustion atmosphere. The introduction of biomass gas alters the chemical reaction atmosphere within the furnace. Hydrogen and carbon monoxide in the biomass gas, as reducing components, can reduce the formed NOx to nitrogen within a specific temperature range. This reduction mechanism fundamentally differs the NOx formation patterns in biomass gas coupled combustion systems from those in pure coal-fired systems. While pure mechanistic models can describe the theoretical laws governing the above physicochemical processes, the inevitable simplification assumptions introduced during model construction, such as the assumption of complete mixing within the reactor and idealized values ​​of reaction kinetic parameters, lead to systematic deviations between the predicted results of the mechanistic models and actual measurements. While pure data-driven models are fast and can learn nonlinear relationships from historical data, they lack constraints on the physical laws of combustion. Under extreme conditions not covered by training data, they may output predictions that violate physical laws, such as predicting that the sum of the concentrations of flue gas components exceeds 100%, or predicting that the concentration of nitrogen oxides increases when the reducing atmosphere is enhanced.

[0080] Step S20 constructs a multi-model fusion prediction architecture, organically combining the full mixed-flow reactor network mechanism model, the Gaussian process regression model, and the physical information neural network. Using the feature recombination matrix `Mreaction` generated in step S10 as a unified input, it achieves a collaborative prediction mechanism where the mechanism model provides a physical baseline, the data model performs bias correction, and physical constraints are embedded to prevent prediction distortion. The full mixed-flow reactor network mechanism model, based on chemical reaction kinetics, simplifies the complex furnace space into several ideal reactors connected in series, embedding the core reaction equations for nitrogen oxide formation and reduction, and outputting theoretical nitrogen oxide concentrations that conform to stoichiometry, thus establishing a physical baseline trend for the entire prediction system. The Gaussian process regression model, as a non-parametric Bayesian method, can not only learn nonlinear biases that the mechanism model cannot capture but also quantify the uncertainty of the prediction results, providing a statistical basis for assessing the reliability of the prediction results. The physical information neural network, by explicitly incorporating mass conservation, energy conservation, and stoichiometry constraints into the loss function, ensures that the neural network follows physical laws while fitting the data, maintaining physical consistency in the prediction results even under unseen operating conditions. The fusion output of the three models not only provides high-precision point predictions, but also generates confidence intervals and feature contribution analysis, enabling operators to obtain reliable nitrogen oxide prediction values ​​and understand the key factors that cause changes in the prediction results, providing an interpretable basis for boiler combustion optimization and adjustment.

[0081] Further, step S20 includes:

[0082] Step S21: Divide the furnace space into three series-connected fully mixed-flow reactors: a main combustion zone, a reduction zone, and a burnout zone. Use the component label vector attributes of pulverized coal energy and the pulverized coal mass flow rate, and the component label vector attributes of biomass gas energy and the biomass gas volume flow rate in the feature recombination matrix as the inlet boundary conditions of the fully mixed-flow reactor in the main combustion zone to construct a fully mixed-flow reactor network mechanism model. Embed reaction kinetic equations into the fully mixed-flow reactor network mechanism model, solve the mass conservation equation and the energy conservation equation simultaneously, and output the theoretical NOx concentration.

[0083] Specifically, a completely mixed flow reactor (CSTR) is an idealized reactor model in chemical reaction engineering. Its core assumptions are that the materials within the reactor are completely mixed, the concentration and temperature are uniform throughout, and the composition of the outlet material is the same as that inside the reactor. Simplifying the boiler furnace into a CSTR network is a modeling strategy that transforms the complex multiphase reaction flow field in three-dimensional space into a one-dimensional series of reactors. This simplification transforms the computational fluid dynamics problem, which originally required solving a system of three-dimensional partial differential equations, into a lumped parameter model problem that requires solving a system of ordinary differential equations. This reduces computational complexity by several orders of magnitude, meeting the speed requirements for real-time prediction.

[0084] The division of the furnace space into three functional zones is based on the distribution patterns of the physicochemical properties of the combustion process. The main combustion zone, located near the burner nozzle, is the core area where fuel mixes with primary and secondary air, resulting in a vigorous combustion reaction. The temperature in this zone reaches its peak within the furnace, and the oxygen concentration rapidly decreases as the combustion reaction proceeds. Thermal nitrogen oxides and fuel-based nitrogen oxides are generated in large quantities in this zone. The reduction zone, located downstream of the main combustion zone and upstream of the burnout air nozzle, is oxygen-deficient due to the near depletion of primary and secondary air. Hydrogen and carbon monoxide from the biomass gas act as reducing agents in this zone, reducing the generated nitrogen oxides to nitrogen. The efficiency of the reduction reaction is closely related to the temperature level and the concentration of reducing gases in this zone. The burnout zone, located downstream of the burnout air nozzle and between the burnout air nozzle and the furnace outlet, sees the injection of burnout air causing a rebound in oxygen concentration. Unburned carbon particles are completely burned in this zone, while some nitrogen intermediates are re-oxidized back to nitrogen oxides. The series connection of the three zones reflects the upward flow sequence of flue gas in the furnace. The material exiting the upstream zone serves as the material entering the downstream zone, forming a step-by-step transfer relationship of matter and energy.

[0085] The boundary conditions of the characteristic recombination matrix Mreaction and the network mechanism model of the fully mixed-flow reactor are connected as follows: The first row vector of the characteristic recombination matrix contains the coal powder component label vector Ccoal and the coal powder mass flow rate mcoal, which are used to determine the elemental flow rates of coal powder entering the main combustion zone CSTR, specifically including the carbon element flow rate, hydrogen element flow rate, oxygen element flow rate, nitrogen element flow rate, and ash flow rate. The elemental flow rate is calculated based on the product of the mass flow rate and the component mass fraction. The second row vector of the characteristic recombination matrix contains the biomass gas component label vector Cgas and the biomass gas volumetric flow rate vgas, which are used to determine the molar flow rates of gaseous combustible components entering the main combustion zone CSTR, specifically including the hydrogen molar flow rate, carbon monoxide molar flow rate, and methane molar flow rate. The molar flow rate is calculated based on the product of the volumetric flow rate and the component volume fraction, divided by the molar volume under standard conditions. In boilers employing staged air combustion (CSTR) technology, the main combustion zone uses a lower-than-theoretical-air distribution strategy to suppress the generation of thermal nitrogen oxides (NOx), resulting in an oxygen-deficient combustion state. Only partial oxidation of the combustible components in the biomass gas occurs, while unreacted hydrogen and carbon monoxide are carried downstream to the reduction zone as flue gas components. The reduction zone maintains an oxygen-deficient atmosphere without supplemental oxygen, allowing residual hydrogen and carbon monoxide from the main combustion zone outlet flue gas to act as reducing agents and react with NOx. In the third row vector of the feature recombination matrix, the furnace operating parameters include secondary air volume to determine the oxygen and nitrogen flow rates entering the main combustion zone (CSTR), burnout air volume to determine the supplemental oxygen flow rate entering the burnout zone (CSTR), and furnace temperature and oxygen flow rate as reference check values ​​for solving the CSTR network. These boundary conditions ensure that the input to the CSTR network mechanism model and the feature recombination matrix output in step S10 are strictly correlated in physical time and space, guaranteeing that the mechanism model calculates the amount of NOx generated related to the fuel actually participating in combustion at the current reaction moment.

[0086] The embedded reaction kinetic equations follow the chemical reaction mechanism of nitrogen oxide formation and reduction. The reaction equations embedded in the main combustion zone CSTR include three categories: the pulverized coal pyrolysis and combustion reaction equations describe the volatile matter release, volatile matter oxidation, and coke heterogeneous combustion processes, with reaction rates expressed in Arrhenius form and rate constants changing exponentially with temperature; the extended Zeldovich mechanism equations describe the formation process of thermal nitrogen oxides, which includes three elementary reactions involving the attack of oxygen atoms on nitrogen molecules, the attack of nitrogen atoms on oxygen molecules, and the reaction of nitrogen atoms on hydroxyl groups, with reaction rates strongly dependent on temperature, and higher temperatures resulting in a greater rate of thermal nitrogen oxide formation; and the fuel nitrogen conversion mechanism equations describe the conversion process of organic nitrogen in pulverized coal and biomass gas into nitrogen oxides and nitrogen. Organic nitrogen is first released as intermediate products such as hydrogen cyanide and ammonia, which are oxidized to nitrogen oxides in an oxidizing atmosphere and converted into nitrogen in a reducing atmosphere. The reduction zone CSTR contains a set of nitrogen oxide reduction reaction equations, describing the process by which hydrogen and carbon monoxide reduce nitrogen oxides to nitrogen. The reaction equations are: NO reacts with H2 to produce nitrogen and water vapor, and NO reacts with CO to produce nitrogen and carbon dioxide. The reaction rate is positively correlated with both the concentration of reducing agent and the concentration of nitrogen oxides.

[0087] The process of solving the mass conservation equation and the energy conservation equation simultaneously is as follows. The mass conservation equation is established separately for each element, including carbon, hydrogen, oxygen, and nitrogen, and for each component, including oxygen, nitrogen, carbon dioxide, water vapor, carbon monoxide, hydrogen, and nitrogen oxides. The equation takes the form: the mass flow rate of a component entering the reactor minus the mass flow rate of that component leaving the reactor equals the mass rate at which that component is consumed or generated due to the chemical reaction within the reactor. The energy conservation equation takes the form: the enthalpy of the material entering the reactor plus the heat released by the chemical reaction equals the enthalpy of the material leaving the reactor plus the heat lost to the furnace wall. The heat dissipation from the furnace wall is estimated based on the furnace structure parameters and measured wall temperature. Since the reaction rate and physical properties are both temperature-dependent, and temperature is determined by the energy conservation equation, there is a coupling relationship between the mass conservation equation and the energy conservation equation, requiring an iterative solution method. The iterative solution process is as follows: Assume an initial temperature value for each CSTR, calculate the reaction rate and component concentration changes at that temperature, substitute them into the energy conservation equation to calculate a new temperature value. If the difference between the new temperature value and the assumed value is less than the convergence tolerance, the iteration ends; otherwise, the new temperature value is used as the assumed value for the next iteration, and the calculation continues until the convergence condition is met. The three CSTRs are solved sequentially in series, with the outlet state of the upstream CSTR serving as the inlet boundary condition for the downstream CSTR. After the solution is completed, the nitrogen oxide concentration is extracted from the outlet component concentration of the burnout zone CSTR; this concentration is the theoretical NOx concentration output by the CSTR network mechanism model.

[0088] The CSTR network mechanism model constructed in step S21 provides a physical benchmark based on chemical reaction kinetics for the entire prediction system. This benchmark serves to: establish the basic trend direction of nitrogen oxide concentration changes with input characteristics, ensuring that the theoretical NOx concentration change direction output by the mechanism model remains consistent with the actual physical process regardless of whether the nitrogen content of pulverized coal or the reduction components of biomass gas increase; clarify the stoichiometric relationships between components, ensuring the mass conservation of elements such as carbon, hydrogen, oxygen, and nitrogen before and after the reaction, preventing errors in the prediction results that violate the law of conservation of mass; and provide estimated values ​​for the temperature and component concentration in each reaction region. While these intermediate state quantities are difficult to measure directly, they are of significant reference value for understanding the formation mechanism of nitrogen oxides and optimizing combustion parameters. The computational complexity of the CSTR network mechanism model is far lower than that of the three-dimensional computational fluid dynamics model, meeting the stringent requirements of real-time online prediction for computational speed. Without the mechanistic model from step S21, the Gaussian process regression model in subsequent step S22 will lack a physical benchmark as input features. Its prediction objective will degenerate from "correcting mechanistic model bias" to "directly predicting nitrogen oxide concentration." This degeneration causes the model to lose prior constraints on physical trends, making it prone to producing predictions that contradict physical laws in operating conditions where training data is sparse. The theoretical NOx concentration output from step S21 provides a reference benchmark for the prediction objective in step S22, enabling the Gaussian process regression model to focus on learning the deviation characteristics between the mechanistic model and actual measurements.

[0089] Step S22: Construct a Gaussian process regression model by combining the feature recombination matrix with the theoretical NOx concentration as the input vector of the Gaussian process regression model; obtain the measured NOx concentration and train the Gaussian process regression model using the difference between the measured NOx concentration and the theoretical NOx concentration as the prediction target; input the input vector of the Gaussian process regression model into the trained Gaussian process regression model and output the mean of the prediction bias and the prediction variance; calculate the corrected NOx concentration based on the mean of the prediction bias.

[0090] Specifically, Gaussian Process Regression (GPR) is a nonparametric regression method based on a Bayesian framework. Its core idea is to treat the objective function to be predicted as a stochastic process following a multivariate Gaussian distribution. By inferring the posterior distribution of the objective function at unobserved points using observed data, it simultaneously obtains the predicted mean and prediction uncertainty. The essential difference between GPR and parametric regression methods, such as linear regression and multinomial regression, lies in the fact that parametric methods assume the objective function belongs to a specific family of functions and estimate a finite number of parameters, while GPR directly models the function space without presupposing the specific form of the function. Therefore, it has stronger nonlinear fitting capabilities and more flexible adaptability. The difference between GPR and other nonparametric methods, such as kernel regression and k-nearest neighbor regression, is that GPR is based on rigorous probability theory and can output the complete probability distribution of the predicted results, not just point estimates. This characteristic naturally gives it the ability to quantify uncertainty.

[0091] The input vector Xgpr of the Gaussian process regression model is constructed as follows: All elements of the feature recombination matrix Mreaction output in step S10 are flattened into a one-dimensional vector. This vector sequentially contains five components of the pulverized coal component label vector Ccoal, the pulverized coal mass flow rate mcoal, the five components of the biomass gas component label vector Cgas, the biomass gas volumetric flow rate vgas, and four components of the furnace operating parameters. The theoretical NOx concentration output in step S21 is appended to the end of the above feature vector to form the complete input vector Xgpr. The design intention of using the theoretical NOx concentration as one of the input features is to enable the Gaussian process regression model to learn the regular changes in the deviation between the predicted and actual values ​​of the mechanistic model, rather than directly predicting the nitrogen oxide concentration from the original features. This design implicitly embeds the physical knowledge of the mechanistic model into the data-driven model learning process.

[0092] The predicted target ΔNOx is defined as the difference between the measured NOx concentration and the theoretical NOx concentration. theory The difference is as follows. The measured NOx concentration was obtained through a flue gas analyzer installed in the tail flue of the boiler. The flue gas analyzer uses chemiluminescence or ultraviolet absorption to measure the mass concentration of nitrogen oxides in the flue gas. The design consideration of setting the prediction target as the deviation value rather than the absolute concentration value is that the range and magnitude of the deviation value are usually smaller than those of the absolute concentration value, making the mapping relationship that the Gaussian process regression model needs to learn simpler, which is conducive to improving the model's fitting accuracy and generalization ability; the deviation value reflects the systematic error caused by the simplification assumptions of the mechanistic model. These errors are often related to specific operating conditions. For example, the deviation of the mechanistic model caused by uneven turbulent mixing in the furnace under high load conditions is different from that under low load conditions. By learning the correlation between the deviation and the operating conditions, the Gaussian process regression model can achieve targeted correction.

[0093] The training process of a Gaussian process regression model includes the following steps. The choice of the covariance function, also known as the kernel function, determines the smoothness and correlation structure of the Gaussian process. In this embodiment, a radial basis function (RBF) kernel is used as the basic covariance function. The RBF kernel is expressed as a negative exponential function of the Euclidean distance between input vectors. Its characteristic is that the closer the input points are, the stronger the correlation between their output values. This characteristic aligns with the physical intuition that similar operating conditions in combustion processes lead to similar deviations. Hyperparameter optimization is the core step in the training process. Hyperparameters include the length scale parameter and the signal variance parameter of the RBF kernel. The length scale parameter controls the distance at which the correlation of the target value significantly decreases after a change in input features. The signal variance parameter controls the typical magnitude of the target value's deviation from the mean. Hyperparameters are optimized by maximizing the marginal likelihood function. The marginal likelihood function measures the probability of observed data occurring under the current hyperparameter settings. The optimization process uses gradient ascent or quasi-Newton methods to iteratively search for hyperparameter combinations that maximize the marginal likelihood function.

[0094] The prediction output of a Gaussian process regression model consists of two parts: the mean prediction bias μΔ and the prediction variance σ. 2 Δ. The mean prediction bias μΔ represents the best estimate of the model's bias against the mechanistic model under the current input conditions. This value, superimposed on the theoretical NOx concentration, yields the corrected predicted NOx concentration. In other words, the corrected NOx concentration equals the sum of the theoretical NOx concentration and the mean prediction bias. Theoretical NOx concentration NOx theory The predicted NOx concentration, calculated by the mechanistic model, reflects the physicochemical laws of the combustion process, but contains systematic biases due to the model's simplification assumptions. The mean prediction bias μΔ is predicted by a Gaussian process regression model after learning the differences between the mechanistic model's predicted and measured values ​​from historical data; it represents the expected bias of the mechanistic model under current operating conditions. Adding the two values ​​yields the data-driven corrected NOx concentration prediction, which retains the physical trends established by the mechanistic model while compensating for its systematic errors. The prediction variance σ... 2 Δ represents the uncertainty of the model regarding the mean prediction deviation μΔ. A larger variance indicates lower confidence in the model's prediction of the current operating condition. The physical meaning of prediction variance can be understood from two aspects: when the input vector Xgpr is far from the sample points in the training data, meaning the current operating condition was not fully covered during the training phase, the prediction variance automatically increases, reflecting the model's lack of confidence in predicting unseen operating conditions; when the dispersion of the target value in the training data is large, meaning the deviation value fluctuates significantly under similar operating conditions, the prediction variance also increases accordingly, reflecting the randomness of the deviation itself. Prediction variance provides necessary statistical information for generating confidence intervals in the subsequent step S24.

[0095] The Gaussian process regression model constructed in step S22 achieves data-driven correction of the systematic bias of the mechanistic model. Compared with traditional bias correction methods, the non-parametric nature of Gaussian process regression allows the model to adaptively capture complex nonlinear relationships without pre-setting the functional form between the bias and input features, avoiding fitting failure or overfitting due to inappropriate selection of the functional form. Its probabilistic prediction characteristics enable the model to simultaneously output the mean and variance of the prediction bias, providing a quantitative indicator for assessing the reliability of the prediction results. Its data efficiency characteristics allow the model to achieve good fitting results on small to medium-sized datasets, exhibiting lower dependence on large-scale data compared to deep neural networks, thus adapting to the relatively limited operational data of biomass gas coupled power generation systems. Without the Gaussian process regression correction in step S22, the theoretical NOx concentration output in step S21 would directly participate in subsequent fusion calculations. Since the systematic bias of the mechanistic model is not corrected, the fusion prediction results will inherit the bias characteristics of the mechanistic model, significantly reducing prediction accuracy in operating conditions with large biases. Step S21 provides a coarse prediction result based on the physical mechanism, and step S22 performs fine correction on this basis. The combination of the two realizes the effective integration of physical prior constraints and data-driven learning, so that the final corrected prediction value not only maintains the physical trend direction established by the mechanism model, but also eliminates the systematic bias caused by the simplification assumptions of the mechanism model.

[0096] Step S23: Construct a physical information neural network, define physical constraint loss and data loss, wherein the physical constraint loss includes mass conservation constraint loss, energy conservation constraint loss and stoichiometry constraint loss; use the weighted sum of data loss and physical constraint loss as the total loss function, train the physical information neural network by minimizing the total loss function, input the feature recombination matrix into the trained physical information neural network, and output the predicted value of the physical information neural network;

[0097] Specifically, Physical Information Neural Networks (PINNs) are deep learning architectures that embed physical laws as soft constraints into the neural network training process. Their core idea is to explicitly add physical equation residuals to the loss function, forcing the neural network to satisfy pre-defined physical laws while fitting observed data. Unlike traditional neural networks that rely solely on data for supervised learning, PINNs utilize physical laws to provide additional regularization constraints. These constraints are particularly important when training data is sparse or noisy, guiding the model to converge to a solution space that conforms to physical laws and avoiding physically impossible predictions.

[0098] The network structure of the physical information neural network is designed as follows: the number of neurons in the input layer is equal to the number of elements in the flattened feature reconstruction matrix Mreaction. The hidden layers adopt a fully connected structure with N hidden layers, each containing N neurons. The activation function between the hidden layers is the hyperbolic tangent function (tanh), whose output range is between -1 and +1, exhibiting smooth differentiability, which is beneficial for the stable backpropagation of gradients. For example, Nhidden can be set to 4, and Nneuron can be set to 128. The output layer contains a single neuron, outputting the predicted nitrogen oxide concentration (NOx) value from the physical information neural network. pinn The output layer does not use an activation function, maintaining a linear output to accommodate the continuous numerical range of nitrogen oxide concentration.

[0099] The physical constraint loss consists of three components: mass conservation constraint loss Lmass, energy conservation constraint loss Lenergy, and stoichiometric constraint loss Lchem. These three are added together to form the total physical constraint loss Lphysics, i.e., Lphysics = Lmass + Lenergy + Lchem. This formula uses a simple summation method to combine the three types of physical constraint losses with equal weight. The design logic is that mass conservation, energy conservation, and stoichiometry are fundamental physical laws that must be simultaneously satisfied in the combustion process. All three are equally important; violation of any one constraint will render the prediction results meaningless. By adding the three types of constraint losses to form the total physical constraint loss, the neural network is forced to simultaneously satisfy all three physical laws during training, ensuring consistency in the predicted output across all physical dimensions. The calculation of the mass conservation constraint loss Lmass is based on the elemental balance relationship between the inlet and outlet materials of the furnace. The mass flow rates of carbon, hydrogen, oxygen, and nitrogen entering the furnace are calculated based on the coal powder composition and flow rate information in the feature reconstruction matrix. The mass flow rates of gaseous carbon, hydrogen, and oxygen entering the furnace are calculated based on the biomass gas composition and flow rate information. These two calculations are then added together to obtain the total inlet element mass flow rate. Based on the nitrogen oxide concentration predicted by the neural network and the concentrations of other flue gas components estimated based on combustion stoichiometry, combined with the furnace outlet flue gas flow rate, the mass flow rates of carbon, hydrogen, oxygen, and nitrogen in the outlet flue gas are calculated. Specifically, the carbon dioxide concentration is estimated based on the product of the inlet carbon flow rate and the degree of combustion completeness; the water vapor concentration is estimated based on the inlet hydrogen flow rate according to the stoichiometry of hydrogen combustion to produce water; the residual oxygen concentration is estimated based on the total inlet oxygen content minus the oxygen consumed during combustion; and the nitrogen concentration is estimated based on the sum of the nitrogen content in the inlet air and the portion of fuel nitrogen converted to nitrogen. Mass conservation requires that the inlet mass flow rate and outlet mass flow rate of each element be equal. Therefore, the mass conservation violation is defined as the sum of the squares of the differences between the inlet and outlet flow rates of each element, and the mass conservation constraint loss Lmass is the average of this violation over all training samples.

[0100] The energy conservation constraint loss Lenergy is calculated based on the balance between the chemical energy of the fuel and the enthalpy of the flue gas. The chemical energy input power of the pulverized coal is calculated based on its industrial analytical composition and mass flow rate. Chemical energy is calculated as the product of the lower heating value (LCV) of the pulverized coal and its mass flow rate; the LCV can be estimated from the industrial analytical composition using empirical formulas. The chemical energy input power of the biomass gas is calculated based on its composition and flow rate. Chemical energy is calculated as the sum of the products of the heat of combustion and molar flow rates of each combustible component, including hydrogen, carbon monoxide, and methane. The total chemical energy input of the fuel equals the sum of the chemical energy of the pulverized coal and the chemical energy of the biomass gas. The enthalpy output power of the flue gas is calculated based on the flue gas flow rate, average flue gas temperature, and specific heat capacity, where the flue gas temperature is obtained from the furnace operating parameters. Energy conservation requires that the chemical energy input of the fuel equals the enthalpy output of the flue gas plus the heat loss from the furnace walls and the loss from incomplete combustion. Therefore, the energy conservation violation is defined as the square of the difference between the chemical energy input and the effective heat output. The energy conservation constraint loss Lenergy is the average of this violation over all training samples.

[0101] The stoichiometric constraint loss Lchem constrains the stoichiometric relationships of specific chemical reactions. Taking the carbon monoxide oxidation reaction as an example, the chemical equation for this reaction is that two moles of carbon monoxide react with one mole of oxygen to produce two moles of carbon dioxide, with a stoichiometric ratio of 2:1:2. The stoichiometric constraint requires that the predicted ratio of carbon monoxide consumption to oxygen consumption equals 2, and that the carbon monoxide consumption equals the carbon dioxide production. The stoichiometric violation is defined as the degree to which the predicted value deviates from the above stoichiometric relationship, and the stoichiometric constraint loss Lchem takes the average of this violation over all training samples.

[0102] Data loss Ldata is calculated using the mean squared error function and is defined as the physical information neural network prediction value NOx. pinn Compared with the measured nitrogen oxide concentration (NOx) measuredThe average of the squared difference over all training samples. Data loss measures the fit between the model's predictions and the observed data, and is the standard loss form for supervised learning. The total loss function Ltotal is defined as the weighted sum of the data loss Ldata and the physical constraint loss Lphysics, i.e., Ltotal = Ldata + λ × Lphysics, where λ is the physical constraint weight coefficient. The physical constraint weight coefficient λ controls the relative importance of physical constraints in the total loss function: if λ is too small, the physical constraint effect is weak, and the model may output predictions that violate physical laws; if λ is too large, the physical constraint is too strong, which may inhibit the model's ability to learn from the real fluctuations in the observed data. The method for determining the physical constraint weight coefficient λ is to compare the model's prediction accuracy and physical violation degree under different λ values ​​on the validation dataset, and select a λ value that makes both acceptable. For example, λ can be a value between 0.1 and 1.0.

[0103] The training of the physical information neural network employs a gradient descent optimization algorithm. During training, the total loss function Ltotal is calculated for each training sample, with respect to the gradients of the network parameters, including the weights and biases of each layer. The network parameters are updated according to the gradient direction to reduce the total loss function value. The optimizer uses the Adam algorithm, which combines the advantages of momentum and adaptive learning rate, exhibiting good convergence performance when handling non-convex optimization problems. The initial learning rate is set to ηinit. During training, a learning rate decay strategy is employed, multiplying the learning rate by a decay factor γdecay every nepoch training epochs. For example, ηinit can be set to 0.001, nepoch to 10, and γdecay to 0.9. The training terminates when the total loss function value changes less than the convergence threshold over a consecutive number of epochs, or when the preset maximum number of training epochs is reached. After training, the feature recombination matrix Mreaction is input into the physical information neural network, and the network outputs the predicted value NOx. pinn Due to the existence of physical constraint loss during training, NOx pinn It not only has high fitting accuracy in the operating conditions covered by the training data, but also maintains adherence to physical laws in extreme operating conditions where the training data is sparse or not covered, avoiding abnormal prediction results that do not conserve mass or energy.

[0104] The physical information neural network constructed in step S23 provides the prediction system with a neural network prediction channel guaranteed by physical constraints. Compared with the Gaussian process regression model in step S22, the advantage of the physical information neural network lies in its guarantee of physical consistency for unseen operating conditions: the Gaussian process regression model has increased prediction variance in regions far from the training data but does not guarantee that the prediction mean conforms to physical laws, while the physical information neural network, through the physical constraint term in the loss function, is forced to output prediction results that satisfy mass and energy conservation even under unseen operating conditions. This physical constraint sets an insurmountable boundary for the neural network, restricting the prediction results to a physically feasible solution space. Without the physical information neural network in step S23, the multi-model fusion prediction will lose the guarantee of physical constraints, and under extreme operating conditions, the correction value output by the Gaussian process regression model may cause the fusion prediction results to violate physical laws, reducing the reliability of the prediction results. Step S23, together with steps S21 and S22, forms a three-channel parallel prediction architecture: the CSTR network mechanism model in step S21 provides theoretical prediction based on chemical kinetics, the Gaussian process regression model in step S22 provides data-driven bias correction, and the physical information neural network in step S23 provides neural network prediction with physical constraints. The three approach the real nitrogen oxide concentration from different perspectives, providing diversified prediction information for the fusion decision in step S24.

[0105] Step S24: The theoretical NOx concentration, the corrected NOx concentration, and the predicted value of the physical information neural network are weighted and fused to generate the final NOx predicted value; the composite standard deviation is calculated based on the prediction variance output by the Gaussian process regression model to generate the confidence interval of the final NOx predicted value.

[0106] Specifically, weighted fusion is a method that sums the predicted outputs of multiple sub-models using weighted coefficients to obtain the final predicted value. Its basic principle is to leverage the complementary advantages of different models and, through reasonable weight allocation, make the fusion result superior to the prediction result of any single model. In step S24, the predicted value involved in the fusion includes three sources: the theoretical NOx concentration output in step S21. theory Representing the baseline prediction based on physical mechanisms, the corrected prediction value output in step S22 is the NOx concentration, specifically the NOx... theory +μΔ represents the prediction after data-driven bias correction, and the physical information neural network prediction value NOx output in step S23. pinn This represents a neural network prediction with physical constraints.

[0107] Final NOx prediction value NOx final The calculation formula for NOx is: final =w1×NOx theory +w2×(NOx) theory+μΔ)+w3×NOx pinn The weighting coefficients w1, w2, and w3 represent the weighting coefficients of the prediction results from the three sub-models, and their sum equals 1 to ensure the unbiasedness of the fusion result. The weighting coefficients are determined using recursive least squares (RLS) with online dynamic updates. The basic idea of ​​RLS is to dynamically adjust the weighting coefficients based on the error between the latest observed measured nitrogen oxide concentration and the prediction values ​​of each sub-model, minimizing the prediction error of the weighted fusion result. This dynamic updating of the weighting coefficients allows the fusion prediction to adapt to changes in the prediction performance of each sub-model under different operating conditions. Under certain operating conditions, the assumptions of the mechanistic model are close to the actual situation, and the prediction error of the theoretical NOx concentration is small. In this case, w1 will automatically increase to increase the weight of the mechanistic model prediction. Under other operating conditions, the mechanistic model has significant biases, while the Gaussian process regression correction effect is significant. In this case, w2 will automatically increase to increase the weight of the corrected prediction. This adaptive mechanism enables the fusion prediction to achieve near-optimal prediction accuracy under various operating conditions without requiring manual intervention in weight setting.

[0108] Composite standard deviation σ total The calculation comprehensively considers the prediction variance σ of the Gaussian process regression model output. 2 Uncertainties introduced by Δ and the fusion process. Prediction variance σ of Gaussian process regression. 2 Δ reflects the statistical uncertainty of the biased prediction, which mainly stems from the limited coverage of the training data and the random fluctuations of the target value itself. The uncertainty introduced by the fusion process arises from the estimation errors of the weight coefficients and the correlation between the prediction errors of each sub-model. The formula for calculating the composite standard deviation is: Where j is the index of each sub-model in the multi-model fusion prediction architecture, Let j be the residual covariance term of the j-th sub-model. Let be the weight coefficient of the j-th sub-model. This represents the weighted sum of the residual covariance terms of each sub-model. For simplified calculations, the residual covariance term can be ignored, and σ can be approximated. total ≈w2× Composite standard deviation σ total The final NOx prediction value is provided. final The overall uncertainty measure.

[0109] Confidence intervals are generated based on the assumption of a normal distribution. Under large sample conditions, the prediction error approximately follows a normal distribution; therefore, confidence intervals can be constructed using the quantiles of the normal distribution. The formula for calculating the 95% confidence interval is: Lower confidence limit = NOx final -z crit ×σ total Confidence ceiling = NOx final +z crit ×σtotal , where z crit This is the critical value for the standard normal distribution, corresponding to the 95% confidence level for z. crit It is approximately equal to 1.96. The physical meaning of a confidence interval is that, under current operating conditions, the actual nitrogen oxide concentration has a 95% probability of falling within this range. Confidence intervals provide operators with an intuitive reference for assessing the reliability of prediction results: a narrow confidence interval indicates high reliability of the prediction results, allowing for combustion optimization adjustments; a wide confidence interval indicates significant uncertainty in the prediction results, requiring cautious acceptance or waiting for more data confirmation.

[0110] For example, setting the current prediction time, the outputs and parameters of each sub-model in the multi-model fusion prediction architecture are as follows: the theoretical NOx concentration output by the fully mixed-flow reactor network mechanism model is 200.0 mg / m³; the mean prediction bias output by the Gaussian process regression model is 10.0 mg / m³, and the prediction variance is 4.0; the physical information neural network output predicts a value of 208.0 mg / m³; the weight coefficients dynamically updated by recursive least squares are as follows: (Corresponding theoretical NOx concentration weight) (Corresponding to the corrected NOx concentration weight). (Corresponding to the weights of the predicted values ​​in the physical information neural network); The calculation process first determines the corrected NOx concentration, which is the sum of the theoretical NOx concentration and the mean of the prediction deviation: mg / m³; then weighted fusion is performed to generate the final NOx prediction value. Substituting the numerical values, we get mg / m³; then calculate the composite standard deviation. Following simplified calculation principles, the composite standard deviation is approximated by multiplying the corrected NOx concentration weighting coefficient by the arithmetic square root of the predicted variance, i.e. mg / m³; finally, the critical value based on the standard normal distribution at a 95% confidence level. Calculate the confidence interval: lower confidence limit mg / m³, upper confidence limit mg cubic meters, i.e., the final NOx forecast The confidence interval is [205.44, 209.36] mg / m³, which indicates that the actual nitrogen oxide concentration under the current operating conditions is... The probability falls within this range.

[0111] Step S20 constructs a multi-model fusion prediction architecture, organically combining the physical rigidity of the mechanistic model, the statistical correction capability of Gaussian process regression, and the physical constraint characteristics of the physical information neural network, achieving high-precision nitrogen oxide prediction that conforms to physical laws. Step S21's CSTR network mechanistic model lays the physical foundation for the entire prediction system. Although its output theoretical NOx concentration has biases due to model simplification, it establishes the correct trend direction of nitrogen oxide concentration changes with input characteristics, ensuring that subsequent correction and fusion processes do not deviate from physical laws. Step S22's Gaussian process regression model performs data-driven correction for systematic biases in the mechanistic model and outputs a prediction uncertainty metric, enabling the prediction system to have self-evaluation capabilities, identify operating conditions with low prediction reliability, and issue early warnings. Step S23's physical information neural network, by forcibly imposing physical constraints during training, ensures that the neural network prediction does not violate the laws of mass and energy conservation under any operating condition, providing physical feasibility guarantees for prediction results under extreme conditions. The weighted fusion strategy in step S24 dynamically adjusts the weights of each sub-model using the recursive least squares method, enabling the fusion prediction to adaptively select the optimal prediction combination under different operating conditions, while simultaneously generating confidence intervals by integrating the uncertainty information of each sub-model.

[0112] The feature recombination matrix Mreaction output in step S10 undergoes spatiotemporal tracking and recombination processing to ensure that the coal powder characteristics, biomass gas characteristics, and furnace operating parameters of the input multi-model fusion prediction architecture strictly correspond to the same reaction event in physical space and time. Without the spatiotemporal alignment processing in step S10, the sub-models in step S20 would make predictions based on spatiotemporally misaligned input features. No matter how sophisticated the model architecture, it would be unable to correct the causal misalignment of the input data itself, and the prediction results would inevitably exhibit a phase deviation that is out of sync with the actual reaction process. The physical spatiotemporal alignment features provided in step S10, combined with the multi-model fusion prediction in step S20, address the spatiotemporal matching problem of which input features from which time periods are needed to predict the nitrogen oxide concentration at which moment, while the latter solves the complex nonlinear mapping problem of how to accurately predict the nitrogen oxide concentration from the correct input features. The synergy between the two enables the entire prediction system to possess both a physically self-consistent input-output correspondence and high-precision nonlinear prediction capabilities.

[0113] The final NOx prediction and confidence interval generated in step S20 provide multi-dimensional decision support information for boiler combustion optimization. The final NOx prediction can be compared with the emission limit in real time. When the prediction is close to the limit, an early warning is issued and combustion parameter adjustments are triggered. The confidence interval provides a risk assessment basis for adjustment decisions. When the confidence interval is narrow, active optimization measures can be taken, while when the confidence interval is wide, a conservative approach or more data collection is required for confirmation. The redundant design of the multi-model fusion architecture also enhances the robustness of the prediction system. When a sub-model fails to predict due to sensor failure or data anomalies, other sub-models can still provide basically reliable prediction results. The weight adaptive mechanism automatically reduces the weight of the malfunctioning sub-model, minimizing the impact on the fusion prediction.

[0114] Example 2:

[0115] This embodiment, based on Embodiment 1, provides a nitrogen oxide prediction system for a biomass gas coupled boiler, such as... Figure 5 As shown, it includes:

[0116] Feature Recombination Module: Used to construct a virtual transport coordinate system, generate energy packages in the virtual transport coordinate system and perform trajectory tracking, obtain the arrival time of the energy packages to the furnace reaction zone based on the trajectory tracking results, and construct a feature recombination matrix aligned with the reaction time based on the arrival time; the energy packages include pulverized coal energy packages and biomass gas energy packages;

[0117] The multi-model prediction module is used to construct a multi-model fusion prediction architecture that includes a fully mixed-flow reactor network mechanism model, a Gaussian process regression model, and a physical information neural network. Based on the feature recombination matrix, it uses the fully mixed-flow reactor network mechanism model in the multi-model fusion prediction architecture to output the theoretical NOx concentration, uses the Gaussian process regression model in the multi-model fusion prediction architecture to correct the deviation of the theoretical NOx concentration, and uses the physical information neural network in the multi-model fusion prediction architecture to perform physical constraint prediction, obtaining the predicted value of the physical information neural network.

[0118] Weighted fusion module: Used to weight and fuse the theoretical NOx concentration, the corrected NOx concentration, and the physical information neural network prediction value to generate the final NOx prediction value.

[0119] Furthermore, in the feature recombination module, the method for trajectory tracking includes:

[0120] The virtual conveyor belts for pulverized coal and biomass gas in the virtual transport coordinate system are spatially discretized to generate spatially discretized micro-elements.

[0121] Trajectory tracking of pulverized coal energy packages and biomass gas energy packages is performed in the spatial discretized infinitesimal elements of the virtual transport coordinate system, and the real-time position coordinates of the pulverized coal energy packages and biomass gas energy packages at each moment are calculated.

[0122] The method for determining whether the energy package has reached the furnace reaction zone includes:

[0123] Calculate the total length of the virtual conveyor belt path for pulverized coal and the virtual conveyor belt path for biomass gas; extract the position coordinates of the pulverized coal energy package and the position coordinates of the biomass gas energy package from the real-time position coordinates.

[0124] If the value corresponding to the position coordinate of the pulverized coal energy package is greater than or equal to the total length of the virtual pulverized coal conveyor path, then it is determined that the pulverized coal energy package has reached the furnace reaction zone.

[0125] If the position coordinates of the biomass gas energy package are greater than or equal to the total length of the virtual biomass gas conveyor belt path, then the biomass gas energy package is determined to have reached the furnace reaction zone.

[0126] The method for constructing the feature recombination matrix includes:

[0127] Based on the current reaction time in the furnace, among all energy packages that have completed trajectory tracking, search for pulverized coal energy packages and biomass gas energy packages whose arrival time matches the current reaction time; extract the component label vector attributes and pulverized coal mass flow rate of the matching pulverized coal energy packages, as well as the component label vector attributes and biomass gas volume flow rate of the matching biomass gas energy packages, and combine them with the furnace operating parameters at the current reaction time to construct a feature recombination matrix aligned with the reaction time.

[0128] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0129] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0130] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting nitrogen oxide emissions from a biomass gas coupled boiler, characterized in that, The method includes: A virtual transport coordinate system is constructed, and energy packages are generated and tracked within the virtual transport coordinate system. The arrival time of the energy packages in the furnace reaction zone is obtained based on the trajectory tracking results. A feature recombination matrix aligned with the reaction time is constructed based on the arrival time. The energy packages include pulverized coal energy packages and biomass gas energy packages. A multi-model fusion prediction architecture was constructed, comprising a fully mixed-flow reactor network mechanism model, a Gaussian process regression model, and a physical information neural network. Based on the feature recombination matrix, the theoretical NOx concentration was output using the fully mixed-flow reactor network mechanism model in the multi-model fusion prediction architecture. The theoretical NOx concentration was then corrected for deviation using the Gaussian process regression model to obtain the corrected NOx concentration. Finally, the physical information neural network was used for physical constraint prediction to obtain the predicted value. The theoretical NOx concentration, the corrected NOx concentration, and the physical information neural network prediction value are weighted and fused to generate the final NOx prediction value.

2. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 1, characterized in that, The method for constructing the virtual transport coordinate system includes: The physical parameters of the pulverized coal conveying path and the biomass gas conveying path are determined. Based on the physical parameters, the pulverized coal conveying path and the biomass gas conveying path are mapped to a virtual pulverized coal conveyor belt and a virtual biomass gas conveyor belt, respectively. A virtual transport coordinate system containing the virtual pulverized coal conveyor belt and the virtual biomass gas conveyor belt is established.

3. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 2, characterized in that, The origin of the pulverized coal virtual conveyor belt is set at the coal feeder outlet, and the endpoint is set at the furnace reaction zone inlet; the origin of the biomass gas virtual conveyor belt is set at the biomass gas valve, and the endpoint is set at the furnace reaction zone inlet.

4. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 3, characterized in that, The method for trajectory tracking includes: The virtual conveyor belts for pulverized coal and biomass gas in the virtual transport coordinate system are spatially discretized to generate spatially discretized micro-elements. Trajectory tracking of pulverized coal energy packages and biomass gas energy packages is performed in the spatial discretized infinitesimal elements of the virtual transport coordinate system, and the real-time position coordinates of the pulverized coal energy packages and biomass gas energy packages at each moment are calculated.

5. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 4, characterized in that, The method for determining whether the energy package has reached the furnace reaction zone includes: Calculate the total length of the virtual conveyor belt path for pulverized coal and the virtual conveyor belt path for biomass gas; extract the position coordinates of the pulverized coal energy package and the position coordinates of the biomass gas energy package from the real-time position coordinates. If the value corresponding to the position coordinate of the pulverized coal energy package is greater than or equal to the total length of the virtual pulverized coal conveyor path, it is determined that the pulverized coal energy package has reached the furnace reaction zone. If the position coordinates of the biomass gas energy package are greater than or equal to the total length of the virtual biomass gas conveyor path, then the biomass gas energy package is determined to have reached the furnace reaction zone.

6. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 5, characterized in that, The method for constructing the feature recombination matrix includes: Based on the current reaction time in the furnace, among all energy packages that have completed trajectory tracking, search for pulverized coal energy packages and biomass gas energy packages whose arrival time matches the current reaction time; extract the component label vector attributes and pulverized coal mass flow rate of the matching pulverized coal energy package, as well as the component label vector attributes and biomass gas volume flow rate of the matching biomass gas energy package, and combine them with the furnace operating parameters at the current reaction time to construct a feature recombination matrix aligned with the reaction time.

7. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 6, characterized in that, The method for constructing the network mechanism model of the fully mixed-flow reactor includes: The furnace space is divided into three series-connected fully mixed-flow reactors: the main combustion zone, the reduction zone, and the burnout zone. The component label vector attributes of pulverized coal energy and the pulverized coal mass flow rate, as well as the component label vector attributes of biomass gas energy and the biomass gas volume flow rate in the feature recombination matrix, are used as the inlet boundary conditions of the fully mixed-flow reactor in the main combustion zone to construct a network mechanism model of the fully mixed-flow reactor.

8. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 7, characterized in that, The input vector of the Gaussian process regression model is composed of a feature recombination matrix and the theoretical NOx concentration; the Gaussian process regression model outputs the mean prediction deviation and the prediction variance, and the mean prediction deviation is added to the theoretical NOx concentration to obtain the corrected NOx concentration.

9. The method for predicting nitrogen oxides in a biomass gas coupled boiler according to claim 8, characterized in that, The physical information neural network is trained by minimizing the total loss function, which is a weighted sum of the defined physical constraint loss and data loss. The physical constraint loss is composed of the sum of the mass conservation constraint loss, the energy conservation constraint loss, and the stoichiometric constraint loss.

10. A nitrogen oxide prediction system for a biomass gas coupled boiler, used to implement the nitrogen oxide prediction method for a biomass gas coupled boiler according to any one of claims 1-9, characterized in that, The system includes: Feature Recombination Module: Used to construct a virtual transport coordinate system, generate energy packages in the virtual transport coordinate system and perform trajectory tracking, obtain the arrival time of the energy packages to the furnace reaction zone based on the trajectory tracking results, and construct a feature recombination matrix aligned with the reaction time based on the arrival time; the energy packages include pulverized coal energy packages and biomass gas energy packages; The multi-model prediction module is used to construct a multi-model fusion prediction architecture that includes a fully mixed-flow reactor network mechanism model, a Gaussian process regression model, and a physical information neural network. Based on the feature recombination matrix, it uses the fully mixed-flow reactor network mechanism model in the multi-model fusion prediction architecture to output the theoretical NOx concentration, uses the Gaussian process regression model in the multi-model fusion prediction architecture to correct the deviation of the theoretical NOx concentration, and uses the physical information neural network in the multi-model fusion prediction architecture to perform physical constraint prediction, obtaining the predicted value of the physical information neural network. Weighted fusion module: Used to weight and fuse the theoretical NOx concentration, the corrected NOx concentration, and the physical information neural network prediction value to generate the final NOx prediction value.

Images