Denitration reactor temperature field self-balancing control method and system based on directional ConvLSTM

Abstract

The application discloses a denitration reactor temperature field self-balancing control method and system based on directionality ConvLSTM, belongs to the SCR denitration technical field, and solves the problem of how to improve the precision of the predicted temperature field distribution of a denitration reactor; the application carries out spatial interpolation reconstruction on sparse temperature measurement data based on a wall constraint anisotropic Kriging interpolation algorithm, constructs a directionality ConvLSTM space-time prediction model, carries out multi-objective model prediction control based on the output of the prediction model, constructs a multi-objective cost function, and optimizes an actuator control variable on line; the application designs a directionality convolution long short-term memory network according to the flue gas flow characteristics; the prediction model captures the space-time correlation and utilizes the dominant direction of the flue gas flow to realize more physical-law-conforming prediction, and uses an edge computing device as a DCS external communication module to realize online robust operation, and the application is particularly suitable for the environmental protection island SCR system of power plants, steel plants, cement plants and the like, and realizes safe, efficient and economic intelligent operation of the SCR system.

Description

Technical Field

[0001] This invention belongs to the field of denitrification process technology, and relates to a method and system for self-equilibrium control of temperature field in a denitrification reactor based on directional ConvLSTM. Background Technology

[0002] Denitrification catalytic technologies, including selective non-catalytic reduction (SNCR) and selective catalytic reduction (SCR), are widely used flue gas denitrification methods in industries such as thermal power, steel, and cement. They are also used in combustion devices such as coal-fired power plants and industrial boilers to control nitrogen oxides (NOx). x The main method of emission reduction is through the injection of a reducing agent to react with NO. x A chemical reaction occurs to produce nitrogen and water, thereby reducing the emission concentration of nitrogen oxides.

[0003] Taking Selective Catalytic Reduction (SCR) technology as an example, the denitrification efficiency of the SCR reactor mainly depends on the temperature field distribution characteristics of the flue gas at the reactor inlet, and is affected by near-wall heat transfer conditions, flow field vortices, and ammonia injection distribution strategies. The denitrification efficiency directly affects the catalyst's reactivity, ammonia slip level, and system operational safety. Existing methods for temperature field control in SCR reactors mainly employ the following approaches: (1) Using Kriging interpolation or computational fluid dynamics (CFD) simulation to predict the temperature field distribution. (2) Using convolutional long short-term memory networks (ConvLSTM), recurrent neural networks (RNN), or convolutional neural networks to predict the spatiotemporal changes of temperature and other variables. (3) Using traditional single-objective model predictive control (MPC) or PID control strategies to regulate the operating state of the SCR reactor. For example, the invention patent with publication number CN105675245A discloses a high-precision Kriging test method for predicting flow field distribution based on measured values. Based on the location of the measuring point and the measured parameter values ​​of the measuring point, the Kriging method is applied to calculate the predicted value and draw the contour map. The predicted value of the unknown area is obtained on the contour map and fed back as output.

[0004] Taking coal-fired power plants as an example, thermal power units need to undertake deep peak-shaving tasks, meaning that the load of new energy sources needs to be rapidly and significantly reduced from over 90% to 30% or even lower within a short period of time, and then quickly rebounded when needed. This wide-range, high-ramp-rate deep peak-shaving condition poses a severe challenge to the stable operation of SCR denitrification systems, and existing technologies have revealed the following insurmountable technical bottlenecks:

[0005] (1) Severe nonlinear spatiotemporal dynamics: During rapid load reduction or increase, the flue gas velocity, total volume, and temperature undergo severe nonlinear changes, resulting in severe distortion and fluctuations in the SCR inlet temperature field distribution. Traditional control methods based on sparse measurement points and PID cannot capture and respond to such severe spatiotemporal dynamics, resulting in significant control lag.

[0006] (2) Risk of low-temperature deactivation and blockage under low load: Under low load conditions such as 30%, the flue gas temperature often drops below the lower limit of the catalyst activity temperature window (usually 300-420°C). This not only leads to a sharp reduction in denitrification efficiency and causes NOx emissions to exceed the standard, but more seriously, the excess unreacted ammonia will combine with SO3 in the flue gas to generate ammonium bisulfate (ABS), which is highly corrosive and sticky, thereby blocking the catalyst channels and the downstream air preheater.

[0007] (3) Poor adaptability to all operating conditions: The dynamic characteristics of the system vary greatly in different load ranges (such as high load, low load, and variable load ramping stage). Fixed control models and parameters cannot adapt to the entire operating range, resulting in high ammonia consumption at high load and easy escape at low load, and poor control robustness. With the profound changes in the global energy structure, especially the large-scale grid connection of intermittent new energy sources such as wind power and solar power, the power system has put forward unprecedented requirements for the flexibility of traditional thermal power units. Summary of the Invention

[0008] The technical problem to be solved by this invention is that the temperature field control of existing denitrification reactors has defects such as control lag, poor adaptability and insufficient robustness under wide load and high dynamic conditions such as deep peak shaving.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0010] A self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM includes the following steps:

[0011] S1, Collect sparse temperature measurement data of the denitrification reactor and perform data preprocessing;

[0012] S2 performs spatial interpolation reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features.

[0013] S3, based on the output of spatial interpolation reconstruction, combines working condition data, coal type information and actuator status data as input features to construct a spatiotemporal prediction model based on directional ConvLSTM. The model is trained with a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration and ammonia escape amount for multiple future steps.

[0014] S4, based on the output of the prediction model, performs multi-objective model predictive control, constructs a multi-objective cost function, and optimizes the control quantity of the actuator online;

[0015] S5 performs actuator delay compensation on the optimized control quantity and issues it, sets rollback conditions, and automatically rolls back to the original control strategy when the rollback conditions are met and continues to collect data.

[0016] Further, S1 includes the following steps:

[0017] S11, collect sparse temperature data of the denitrification reactor, including temperature, valve position, load, total flue gas volume, outlet NOx concentration, outlet NH3 concentration and outlet SO2 concentration;

[0018] S12 performs data preprocessing on the collected sparse temperature measurement data, including data time alignment, data cleaning, and data normalization.

[0019] Further, S2 includes the following steps:

[0020] S21, Define the wall grid and calibrate the coordinates of the measuring points;

[0021] S22, spatial interpolation reconstruction of sparse temperature measurement data based on the anisotropic Kriging interpolation algorithm, specifically:

[0022] Constructing anisotropic covariance functions As a kernel function, it is represented using the following logical expression:

[0023]

[0024] In the formula, Let be the spatial displacement vector between two points. Indicates to Transpose operation; The signal variance characterizes the overall fluctuation amplitude of the temperature field; To measure the noise variance, Indicates exponentiation. For the Dirac function, An anisotropic metric tensor, used to define the shape and orientation of spatial correlation, is represented by the following logic. :

[0025]

[0026] In the formula, The relevant lengths along the principal axis, It is a rotation matrix, dominated by the flue gas direction angle. This is determined to align the principal axis of the covariance ellipsoid with the direction of flue gas flow.

[0027] Kriging estimation can be expressed using the following logic:

[0028]

[0029] In the formula, Let x be the temperature value to be predicted at location point x. A column vector consisting of temperature observations from all N actual temperature measurement points. , Given the covariance matrix between the measurement points, Points to be interpolated The covariance vector between the known measurement points and the total covariance vector.

[0030] S23, Set virtual wall points and apply boundary gradient constraints;

[0031] The specific steps of setting virtual wall points are as follows: placing virtual wall points at preset intervals along the wall surface, and setting the virtual wall surface temperature. And the uncertainty of virtual wall temperature ;in, This represents the variance of the actual measured temperature values ​​in the near-wall region. for The corresponding weighting coefficients;

[0032] The boundary gradient constraint specifically refers to: reconstructing the gradient of the temperature field along the wall normal direction. No more than one maximum allowed value This can be represented using the following logic:

[0033]

[0034] In the formula, Represents the temperature field. Indicates the direction of the wall normal. Indicates the boundary gradient constraint value. This indicates the preset heat transfer limit value;

[0035] S24, update the hyperparameters of the Kriging interpolation algorithm online using maximum likelihood estimation. Record hyperparameters The log-likelihood function is obtained by taking the logarithm of the maximum likelihood estimation function. The log-likelihood function is then expressed using the following logic. :

[0036]

[0037] In the formula, Representing the covariance matrix, exponential smoothing filtering is used to smooth the hyperparameters. Periodic online updates are performed using the following logical representation:

[0038]

[0039] In the formula, These are the hyperparameters updated at time t. This is the maximum likelihood estimate at the current moment. This is the smoothing coefficient.

[0040] Further, S3 includes the following steps:

[0041] S31, Define the control vector Coal type indicator vector and working condition vector The input tensor of the prediction model is constructed by combining the output of step S2. ,remember ;in, , , The output of step S2 consists of a three-channel image sequence comprising the temperature field, boundary gradient field, and near-wall mask of the past K frames; the input tensor is... Total number of input channels Recorded as Where K represents the number of historical frames, Indicates the vector dimension;

[0042] S32, Construct a spatiotemporal prediction model based on directional ConvLSTM;

[0043] The prediction model includes a shared encoder and a multi-task decoder. The shared encoder is composed of stacked L layers of directional ConvLSTM units. The multi-task decoder includes a temperature prediction head and an emission prediction head. The temperature prediction head includes several deconvolutional layers or upsampling layers to decode the encoded spatiotemporal features into a temperature field sequence for the next few steps. The emission prediction head includes a global pooling layer and a fully connected layer to regress and predict future NOx concentrations and NH3 escape amounts.

[0044] S33, during the model training phase, uses a boundary-first masking strategy to handle missing tests and fault points, specifically by using a masking strategy with a higher probability than the core region. Higher probability Randomly select the near-wall region of the input tensor. The block is masked by setting all feature values ​​within the block to 0, forcing the prediction model to rely on contextual information and boundary features to infer the value of the missing point;

[0045] S34, construct a sub-model library based on operating condition clustering, and switch or fine-tune the optimal sub-model online according to the current operating status and prediction error; the construction of the sub-model library based on operating condition clustering specifically involves:

[0046] First, clustering algorithms are used to cluster the feature vectors of operating conditions in historical operating data, and similar operating conditions are grouped into the same cluster.

[0047] Then, a dedicated directional ConvLSTM sub-model is trained separately for each load case cluster, forming a sub-model library. During online runtime, the maximum posterior probability is calculated. , choose to The sub-model with the highest probability To make a prediction, the following logical representation can be used:

[0048]

[0049] In the formula, Indicates proportional to, Indicates the current input Next choice The posterior probability of each sub-model This indicates the current operating condition characteristics and the j-th cluster center. The distance between them This represents the center vector of the j-th working condition cluster. Adjustable weights are used to balance the similarity of operating conditions and the accuracy of recent predictions; Let be the root mean square error of the j-th sub-model's predictions on recent historical data.

[0050] Furthermore, in step S32, directional weights are introduced during the update process of each directional ConvLSTM unit. , This can be represented using the following logic:

[0051]

[0052] In the formula, This represents the hidden state at time t. and These represent one-dimensional convolution operations along the x and y directions, respectively, where b is the bias term and the directional weights. , By setting up a small gating network Dynamically generated based on the current operating conditions, using the following logical representation:

[0053]

[0054] In the formula, For the Sigmoid function, The dominant direction of flue gas.

[0055] Furthermore, the prediction model in S32 employs a multi-task loss function based on uncertainty, and the multi-task loss function is represented by the following logic. :

[0056]

[0057] In the formula, The predicted losses are for temperature, NOx, and NH3, respectively. These are the uncertainty terms for temperature, NOx, and NH3, respectively.

[0058] Furthermore, the multi-objective cost function is represented in S4 using the following logic. :

[0059]

[0060] In the formula, Indicates the temperature field tracking term. , , , , These are the weight coefficients for each optimization objective. Let represent the predicted temperature field at time τ. For reference or target temperature field; This represents the boundary gradient penalty term; Indicates performance loss. Represents the performance loss function; This indicates the number of prediction steps for the temperature field. Indicates the control step size. This indicates the control of incremental penalty terms. Indicates control increment; This indicates the penalty term for the control amplitude. The control amplitude is represented; the temperature field efficiency loss function is represented using the following logic. :

[0061]

[0062] In the formula, This means taking the average over all i grid points in the space. For the first Regional weights of spatial grid points This represents the temperature gradient at that point. The optimal reaction temperature for SCR is... is the regularization coefficient.

[0063] Furthermore, the aforementioned Adjustments are made by introducing an online adaptive update law for the weights, using the following logical representation:

[0064]

[0065] In the formula, for Spatial region weights at time points, For the clipping function, through The function limits the updated weights to a preset safety range. Inside, These represent the minimum and maximum values ​​of the spatial region weights, respectively. Forgetting factor, For learning rate, It is the first Several performance metrics.

[0066] Furthermore, the multi-objective cost function in S4 The optimization solution applies constraints, including hard constraints and soft constraints; the hard constraints specifically mean that the magnitude and increment of the control quantity are strictly limited to the physically permissible range, which is represented by the following logic:

[0067]

[0068]

[0069] In the formula, , These represent the maximum and minimum permissible values ​​of the control amplitude, respectively. , These represent the maximum and minimum allowable values ​​for the control quantity increment, respectively;

[0070] The soft constraints specifically involve constraining key output variables such as predicted ammonia slip and NOx emissions, and are represented by the following logic:

[0071]

[0072]

[0073] In the formula, This represents the predicted ammonia release at time t+τ. This indicates the maximum permissible amount of ammonia release. This represents the predicted NOx concentration at the export point at time t+τ. The emission limit for NOx concentration at the export site.

[0074] This invention also provides a self-balancing control system for the temperature field of a denitrification reactor based on directional ConvLSTM, comprising:

[0075] The data processing module is used to collect sparse temperature measurement data from the denitrification reactor and perform data preprocessing.

[0076] The interpolation and reconstruction module is used to perform spatial interpolation and reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features.

[0077] The model prediction module combines the output of spatial interpolation reconstruction with working condition data, coal type information and actuator status data as input features to construct a spatiotemporal prediction model based on directional ConvLSTM. The model is trained with a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration and ammonia escape amount for multiple future steps.

[0078] The MPC module performs multi-objective model predictive control based on the output of the predictive model, constructs a multi-objective cost function, and optimizes the control quantity of the actuator online.

[0079] The control execution module is used to compensate for the delay of the optimized control quantity and issue it, set the rollback conditions, and automatically roll back to the original control strategy when the rollback conditions are met and continue to collect data.

[0080] The advantages of this invention are:

[0081] (1) This invention considers the actual temperature field under wide load conditions. By fusing virtual wall points and boundary gradient constraints, an anisotropic Kriging interpolation algorithm based on wall constraints is constructed. The temperature at unknown locations is predicted based on the correlation between the collected sparse temperature measurement data. The high-resolution temperature field is reconstructed online, and the boundary gradient field and near-wall region mask features are generated simultaneously. Furthermore, the hyperparameters of the interpolation algorithm are updated online adaptively through maximum likelihood estimation to ensure that the reconstructed temperature field can maintain good accuracy and robustness under wide load conditions.

[0082] (2) The present invention designs a directional convolutional long short-term memory network based on the characteristics of flue gas flow. This prediction model deeply couples macroscopic fluid dynamics prior knowledge with microscopic neural network calculation. By capturing spatiotemporal correlation, understanding and utilizing the dominant direction of flue gas flow, it achieves predictions that are more in line with physical laws. When predicting key dynamics such as the drift and evolution of hotspot areas, it effectively improves the accuracy and physical consistency of the prediction model.

[0083] Furthermore, by employing a training loss that balances spatial correlation and anomaly robustness, near-wall block masks and random point masks that closely resemble field failure modes are constructed to improve resistance to time-varying measurement loss and the robustness of the prediction model in the event of sensor failure. Based on the operating condition clustering sub-model library and online switching mechanism, it can effectively cope with the nonlinear dynamics caused by drastic load changes under wide load conditions, and achieve faster and smoother control transition.

[0084] (3) This invention constructs a multi-objective cost function with the optimization objectives of minimizing temperature field tracking error, boundary gradient suppression, reaction efficiency loss, control quantity change and control quantity amplitude, and introduces an economic penalty term to achieve comprehensive optimization effect; introduces the online adaptive update law of weights so that the weights of the MPC controller can be adaptively adjusted to adapt to wide load conditions; and sets multiple constraints to constrain the control amplitude, rate of change, emission environmental protection indicators and economy of the actuator, so that the control command finally output to the actuator can achieve temperature field self-equilibrium, ensure denitrification efficiency and meet economic and environmental protection indicators.

[0085] (4) This invention makes specific engineering innovations and coupling designs for the temperature field distribution characteristics of SCR denitrification reactors under wide load conditions, improves the prediction accuracy and control effect of the prediction model under sparse measurement points, strong boundary gradients and wide load disturbances, and uses edge computing devices as external communication modules of DCS to achieve online robust operation. It is particularly suitable for SCR systems in environmental protection islands of multiple industries such as thermal power, steel, and cement. It can realize the safe, efficient and economical intelligent operation of SCR system without changing the original control logic, and has broad promotion prospects and significant industrial practical value.

[0086] (5) This invention can be widely applied to industrial boilers (industries that convert heat energy by using fuel combustion, such as thermal power generation, gas power generation, biomass power generation, and waste incineration power generation), industrial kilns (including but not limited to glass, cement, ceramics, building materials, and other industries that achieve high-temperature processes such as calcination, smelting, and sintering of materials by fuel combustion or electric heating), sintering machines, blast furnaces, pelletizing, coking, chemical product manufacturing, marine power systems, metallurgy, VOC treatment industries, and other environmental island flue gas treatment scenarios that contain one or more of the following processes: desulfurization, denitrification, and dust removal.

[0087] The flue gas treatment scenarios include, but are not limited to, desulfurization processes, such as dry desulfurization, wet desulfurization, and semi-dry desulfurization (the desulfurizing agents include, but are not limited to, calcium-based desulfurizing agents, sodium-based desulfurizing agents, magnesium-based desulfurizing agents, etc.); denitrification processes, such as selective non-catalytic reduction (SNCR), selective catalytic reduction (SCR), PNCR, ammonia denitrification, and ammonia-free denitrification; and dust removal processes, such as electrostatic precipitators, bag filters, and electrostatic-bag composite dust collectors. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of the signal link of the self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM in Embodiment 1 of the present invention.

[0089] Figure 2 This is a schematic diagram of the spatial interpolation reconstruction based on the anisotropic Kriging interpolation algorithm in Embodiment 1 of the present invention;

[0090] Figure 3 This is a schematic diagram of the structure of the directional ConvLSTM-based spatiotemporal prediction model according to Embodiment 1 of the present invention;

[0091] Figure 4 This is a schematic diagram of the directional ConvLSTM unit in Embodiment 1 of the present invention;

[0092] Figure 5 This is a schematic diagram of the execution process of multi-objective model predictive control according to Embodiment 1 of the present invention. Detailed Implementation

[0093] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0094] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0095] Example 1

[0096] like Figures 1-5 As shown, specifically, this invention uses an SCR denitrification reactor as an example for illustration. This invention discloses a self-balancing control method for a denitrification reactor under wide load conditions based on directional ConvLSTM prediction, including the following steps:

[0097] S1. Collect sparse temperature measurement data from the SCR reactor and perform data preprocessing, including the following steps:

[0098] S11 collects sparse temperature data of the SCR reactor, including temperature, valve position, load, total flue gas volume, outlet NOx concentration, outlet NH3 concentration, and outlet SO2 concentration.

[0099] In this embodiment, process quantities including temperature, valve position, load, and total flue gas volume are collected in real time by the DCS distributed control system, and the NOx, NH3, and SO2 concentrations at the SCR reactor outlet are collected by the CEMS continuous flue gas monitoring system.

[0100] Furthermore, the sparse temperature measurement data also includes coal batch and industrial analysis data, such as ash content, sulfur, volatile matter and lower heating value data, which are obtained through the FMIS fuel management information system.

[0101] S12 performs data preprocessing on the collected sparse temperature measurement data, including data time alignment, data cleaning, and data normalization.

[0102] In this embodiment, all data is first aligned using high-frequency DCS data as the reference timestamp (control cycle of 250ms) and low-frequency data (such as data collected by the CEMS continuous flue gas monitoring system) through forward padding or linear interpolation. Then, box plots or the 3σ principle are used to remove obviously abnormal data points, and sensor drift is periodically calibrated. Finally, all data is scaled to a similar numerical range to serve as the dataset for model training. This embodiment can select... or As the range for data normalization.

[0103] S2 performs spatial interpolation reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features.

[0104] One of the core aspects of this embodiment lies in how to accurately reconstruct the high-resolution temperature field of the entire SCR reactor cross-section based on limited and sparse temperature measurement data. This step is achieved using a spatial interpolation algorithm. To provide comprehensive technical protection, this invention covers a variety of feasible interpolation algorithms and identifies the optimal technical solution. Specifically, step S2 includes the following steps:

[0105] S21, Define the wall mesh and calibrate the coordinates of the measuring points, specifically including the following:

[0106] First, the temperature field matrix resolution of the reconstructed SCR reactor is defined as N×M. The temperature field matrix resolution can be 16×16 or 32×32, or it can be a non-uniform grid customized according to the cross-sectional shape of the SCR reactor.

[0107] Secondly, the area ≤0.3m from the wall of the SCR reactor is defined as the near-wall area, which serves as the boundary zone of the SCR reactor and is a high-incidence area for temperature fluctuations, catalyst wear, and dust accumulation.

[0108] Then, thermocouple arrays or infrared thermal imagers are used to acquire on-site temperature measurement point data, and the coordinates of the measurement points in the grid coordinate system are calculated based on the installation location of the temperature measurement points.

[0109] S22 uses a spatial interpolation algorithm to reconstruct sparse temperature measurement data through spatial interpolation.

[0110] In step S2, the spatial interpolation algorithm can be selected from any one of the inverse distance weighted interpolation method, radial basis function interpolation method, or improved Kriging interpolation method to perform spatial interpolation reconstruction of sparse temperature measurement data.

[0111] In a preferred embodiment, to achieve the highest prediction accuracy and incorporate prior physical knowledge such as flue gas flow direction into the model, this invention introduces an improved anisotropic Kriging interpolation algorithm. This method not only considers distance but also analyzes the spatial structure and correlation of the data through a variogram, providing optimal linear unbiased estimation and quantifying prediction uncertainty. Specifically, the anisotropic Kriging interpolation algorithm includes the following:

[0112] In this embodiment, the core of reconstructing the SCR reactor temperature field based on improved Kriging interpolation is the optimal linear unbiased estimation based on Gaussian process regression. First, an anisotropic covariance function needs to be constructed as the kernel function. The anisotropic covariance function is represented by the following logic. :

[0113]

[0114] In the formula, Let be the spatial displacement vector between two points. Indicates to Transpose operation; The signal variance characterizes the overall fluctuation amplitude of the temperature field; To measure the noise variance; Indicates exponentiation; is the Dirac function, which has a value of 1 when h=0 and 0 otherwise. It is used to represent the variance of the measurement data itself, which is the noise variance. An anisotropic metric tensor, used to define the shape and orientation of spatial correlation, is represented by the following logic. :

[0115]

[0116] In the formula, The relevant lengths along the principal axis; It is a rotation matrix, dominated by the flue gas direction angle. The angle of the dominant flue gas direction is determined to align the principal axis of the covariance ellipsoid with the flue gas flow direction in this embodiment. It can be obtained offline from CFD priors or through online flow field analysis.

[0117] In this embodiment, although the temperature field grid defined in S21 is a two-dimensional N×M grid, the actual flue gas flow is a three-dimensional physical process. This embodiment introduces relevant lengths... This is to enable the model to characterize the spatial correlation in the direction perpendicular to the two-dimensional cross-section, thereby more accurately describing the impact of three-dimensional flow effects such as turbulence on temperature distribution. In practical applications, The value of can be set using prior knowledge from CFD simulation, or used as a hyperparameter for offline optimization. If simplified to a purely two-dimensional problem, it can be... Set it to a large constant so that its effect is minimized.

[0118] Furthermore, the Kriging estimate can be represented using the following logic:

[0119]

[0120] In the formula, This represents the temperature value to be predicted at location point x; A column vector consisting of temperature observations from all N actual temperature measurement points. ; Given the covariance matrix between the measurement points, The point to be interpolated The covariance vector between x and all known measurement points. This formula is the core of the Kriging interpolation algorithm, used to predict the temperature value at any unknown point x. This enables the reconstruction of the entire temperature field. The improved anisotropic Kriging interpolation algorithm obtains the predicted value by weighted summation of the known measurement point temperatures. The weights are calculated by spatial correlation determined by the covariance function, which can achieve the best linear unbiased estimation of the temperature of unknown points.

[0121] In a preferred embodiment, the spatial interpolation algorithm can employ inverse distance-weighted interpolation (IDW). This method is a simple and commonly used spatial interpolation method. Its core idea is that the value of the point to be interpolated is a weighted average of the values ​​of known measurement points around it, with the weights proportional to the inverse of the distance, expressed logically as follows:

[0122]

[0123] In the formula, Let x be the predicted temperature of the point to be interpolated; Let i be the temperature of the i-th known temperature measurement point. For temperature The corresponding weights , Let x be a point and the i-th known point x. i The Euclidean distance between them, where p is a power parameter.

[0124] In a preferred embodiment, the spatial interpolation algorithm can employ radial basis function interpolation (RBF). RBF is an accurate interpolator capable of generating a smooth interpolation surface, representing the values ​​of the points to be interpolated as a linear combination of a set of distance-dependent radial basis functions and their corresponding weights, using the following logical representation:

[0125]

[0126] In the formula, It is a radial basis function, which can be represented by a Gaussian function or a multiquadrant function; Represents the relationship between point x and center point x. i The distance between them; These are the weight coefficients to be solved. All weights can be obtained by solving a system of linear equations. The RBF method has high accuracy and can produce smooth surfaces, making it an effective alternative interpolation method to the Kriging method.

[0127] S23, set virtual wall points and apply boundary gradient constraints.

[0128] In this embodiment, virtual wall points are placed at preset intervals along the wall surface, and the virtual wall temperature is set. And the uncertainty of virtual wall temperature ;in, This represents the variance of the actual measured temperature values ​​in the near-wall region. for The corresponding weighting coefficients, The preset interval can be 0.2~0.3m. In this embodiment, the virtual wall temperature... With weight The calibration can be obtained by joint inversion using the least squares method or Bayesian method, based on measured data from wall thermocouples or infrared thermometers, or by combining CFD simulations with offline thermal balance calculations.

[0129] In this embodiment, virtual wall temperature is introduced. and uncertainty The purpose is to apply physical soft constraints. This is used to provide prior information on boundary temperatures for reconstruction, while This is used to adjust the influence of the prior information in the Kriging interpolation system, when When the value is small, the reconstructed temperature field will more closely approximate the set wall temperature; conversely, a larger value indicates a weaker constraint. By setting virtual avoidance points and boundary gradient constraints, the reconstructed temperature field in the near-wall region can better conform to the actual heat transfer physics, thus avoiding unrealistic temperature interpolation.

[0130] In this embodiment, to make the reconstructed temperature field more consistent with physical reality, a soft constraint is applied during the spatial interpolation reconstruction process. Specifically, this requires the gradient of the reconstructed temperature field along the wall normal direction to be... No more than one maximum allowed value To prevent unrealistically steep temperature changes near the wall, the following logic is used:

[0131]

[0132] In the formula, Represents the temperature field. Indicates the direction of the wall normal. Indicates the boundary gradient constraint value. This indicates the preset heat exchange limit value.

[0133] S24, update the hyperparameters of the Kriging interpolation algorithm online using maximum likelihood estimation. .

[0134] In this embodiment, hyperparameters are recorded. hyperparameters The initial values ​​were obtained through offline calculations of maximum likelihood estimation (MLE) on a large amount of historical temperature field data. Specifically, a maximum likelihood estimation function was constructed, and the log-likelihood function was obtained by taking the logarithm of the maximum likelihood estimation function. The log-likelihood function was then expressed using the following logic. :

[0135]

[0136] In the formula, This represents the covariance matrix, whose elements depend on the hyperparameters. .

[0137] Furthermore, in order to make the hyperparameters To adapt to wide load variations, exponential smoothing filtering is used to adjust hyperparameters. Periodic online updates are performed using the following logical representation:

[0138]

[0139] In the formula, These are the hyperparameters updated at time t. This is the maximum likelihood estimate at the current moment. As the smoothing coefficient, this embodiment takes... The parameters are updated online every 5 to 15 minutes, which can be selected according to actual needs and is not limited to the scope of this embodiment.

[0140] In this embodiment, after reconstructing the temperature field of the SCR reactor based on step S2, a multi-channel feature map is finally output. The multi-channel feature map includes the temperature field T and the boundary gradient. and near-wall mask Wherein, the temperature field T is the complete temperature matrix T obtained by spatial interpolation and reconstruction of the N×M wall mesh by S22; boundary gradient Specifically, the gradient operator is applied to the reconstructed temperature field T to calculate the overall temperature gradient, retaining only the gradient magnitude in the near-wall region and setting it to zero in other regions to obtain the boundary gradient feature map; near-wall mask. Specifically, an N×M binary matrix is ​​created, with the matrix elements corresponding to the near-wall region set to 1 and the elements of the remaining core region set to 0 to generate the near-wall mask image.

[0141] S3, based on the output of spatial interpolation reconstruction, combines operating condition data, coal type information and actuator status data as input features to construct a spatiotemporal prediction model based on directional ConvLSTM. The model is trained with a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration and ammonia emission in the future multiple steps.

[0142] In this embodiment, to accurately predict the spatiotemporal evolution of the temperature field and key emissions within the SCR reactor, a directional convolutional long short-term memory network (directional ConvLSTM) specifically designed for flue gas flow characteristics is proposed. This prediction model achieves more physically consistent predictions by capturing spatiotemporal correlations and understanding and utilizing the dominant direction of flue gas flow. Specifically, S3 includes the following steps:

[0143] S31, Define the control vector Coal type indicator vector and working condition vector The input tensor of the prediction model is constructed by combining the output of step S2. .

[0144] In this embodiment, a control vector is defined. Coal type indicator vector and working condition vector The control vector u t This represents the status data of the actuators, including the status of all adjustable actuators, such as the total ammonia injection volume in each zone, the opening degree of each ammonia injection grid, the flow distribution coefficient of the injection gun, and the angle of the guide baffle, etc.; the coal type indicator vector The information represents coal type, specifically one-hot or embedded vectors, used to adapt the model to different coal sources and blends; the operating condition vectors The operating data represents scalar values ​​that describe the system's operating status, including global operating characteristics such as boiler load, load change rate, total flue gas volume, primary air volume, inlet flue gas temperature, and O2 concentration.

[0145] In this embodiment, the input tensor Xt of the prediction model at time t is a multi-channel composite tensor, represented by the following logic:

[0146]

[0147] In the formula, , , The output of step S2 consists of a three-channel image sequence comprising the temperature field, boundary gradient field, and near-wall mask of the past K frames; the input tensor is... Total number of input channels Recorded as Where K represents the number of historical frames, Represents the vector dimension.

[0148] S32, Construct a spatiotemporal prediction model based on directional ConvLSTM, including the following:

[0149] like Figure 3 As shown, the prediction model includes a shared encoder and a multi-task decoder; the shared encoder is composed of L layers of stacked directional ConvLSTM units. To improve the nonlinear expressive power and training efficiency of the directional ConvLSTM units, the number of hidden layer channels Ch of each directional ConvLSTM unit is set to 32~64. The gating activation function of the directional ConvLSTM units can be SiLU or GLU, such as... Figure 4 As shown.

[0150] Compared to traditional ConvLSTM units that use isotropic convolution kernels and cannot distinguish spatial directions, this embodiment introduces a directional dynamic convolution mechanism, which can dynamically adjust the flow of information in different spatial directions. Specifically, for the update process of each ConvLSTM unit, directional weights are introduced. , This can be represented using the following logic:

[0151]

[0152] In the formula, This represents the hidden state at time t. and represents one-dimensional convolution operations along the x and y directions, respectively, with b being the bias term. In actual computation, it also includes complete gating and update logic consistent with standard ConvLSTM units, such as input gates, forget gates, and output gates; directional weights. , By setting up a small gating network Dynamically generated based on the current operating conditions, using the following logical representation:

[0153]

[0154] In the formula, For the Sigmoid function, In this embodiment, the dominant direction of flue gas is... It can be obtained from CFD prior knowledge or real-time flow field analysis.

[0155] The core advantage of this directional ConvLSTM cell design lies in its deep coupling of macroscopic fluid dynamics prior knowledge (dominant flue gas direction) with microscopic neural network computation. This transforms the prediction model from a mere "black box" that simply fits the data into an intelligent agent capable of "understanding" the physical process. For example, when the flue gas direction changes, the model can automatically allocate more attention to feature extraction in the downstream direction, thus achieving far greater accuracy and physical consistency than traditional models when predicting key dynamics such as the drift and evolution of hotspot regions.

[0156] In this embodiment, a network architecture based on a shared encoder and a task-specific decoder is used, enabling it to simultaneously output multiple prediction results, including the future temperature field, estimated NOx concentration at the outlet, and ammonia emission. This embodiment introduces separable convolutional kernels and gated activation functions along the main flue gas flow direction and laterally. Based on the flue gas flow direction angle obtained in real-time or offline, the combined weights of these two convolutional kernels are dynamically adjusted, allowing the convolutional operation to adapt to changes in the flow field direction. When the flue gas flow is primarily vertical, the prediction model emphasizes the vertical convolutional kernels for feature extraction, and vice versa, achieving deep coupling between the model and the physical process and enhancing the model's sensitivity to boundaries and strip structures. Specifically, the boundary refers to the region with a severe temperature gradient near the reactor wall, while the strip structure refers to long, strip-shaped high- and low-temperature regions in the temperature field caused by uneven ammonia injection or flow field characteristics.

[0157] Furthermore, to enhance the network's attention to boundary effects, this embodiment assigns higher weights to the prediction error of the near-wall region in the loss function by adding a coefficient greater than 1 to the input near-wall region mask and boundary gradient features. That is, when calculating the loss, the error term of the near-wall region is multiplied by a coefficient greater than 1, thereby forcing the model to pay more attention to the prediction accuracy of this region during training.

[0158] In this embodiment, the decoder of the prediction model adopts a multi-task learning framework and connects different prediction heads after the last layer feature map. The temperature prediction head directly performs pixel-level regression on the feature map output by the last layer to predict the temperature field in the future. The emission prediction head performs global average pooling on the feature map to compress the spatial information into a feature vector, and then uses a multilayer perceptron (MLP) regression to predict the total NOx concentration and NH3 escape in the downstream.

[0159] Specifically, the multi-task decoder includes a temperature prediction head and an emission prediction head. The temperature prediction head includes several deconvolutional layers or upsampling layers to decode the encoded spatiotemporal features into a temperature field sequence for the next few steps. The emission prediction head includes a global pooling layer and a fully connected layer to regress and predict future NOx concentrations and NH3 escape amounts.

[0160] Furthermore, to balance the learning difficulty of different tasks, the prediction model adopts a multi-task loss function based on uncertainty, allowing the model to automatically learn the importance of different tasks and represent it through weights. The multi-task loss function is represented by the following logic. :

[0161]

[0162] In the formula, The predicted losses are for temperature, NOx, and NH3, respectively. These are the uncertainties for temperature, NOx, and NH3, respectively. They are learnable parameters that can automatically assign smaller weights to noisy and difficult-to-learn tasks, thereby balancing the learning speed of different tasks and making the training process more stable and efficient.

[0163] Furthermore, in this embodiment, the directional ConvLSTM spatiotemporal prediction model is deployed on an edge computing device as an external communication module of the DCS system. It communicates with the DCS system via OPC UA / Modbus and uses FP16 quantization and operator fusion to accelerate inference.

[0164] S33 uses a boundary-first masking strategy to handle missing tests and fault points during the model training phase.

[0165] In this embodiment, to enhance the robustness of the prediction model to fluctuations in the boundary region, a boundary block mask is used for model training. Specifically, the model is trained using a mask with a probability greater than that of the core region. (e.g., 0.1) a higher probability (e.g., 0.3) Randomly select the near-wall region of the input tensor The block is masked by setting all feature values ​​within the block to 0, forcing the prediction model to rely on contextual information and boundary features to infer the value of the missing point.

[0166] Furthermore, this embodiment employs the Adam optimizer, with an initial learning rate set to... Furthermore, it incorporates a cosine annealing learning rate scheduling strategy to achieve better convergence.

[0167] This embodiment constructs near-wall block masks and random point masks that closely resemble the field failure modes by employing a training loss that balances spatial correlation and anomaly robustness, thereby improving the resistance to time-varying missing measurements and the robustness of the prediction model in the event of sensor failure.

[0168] S34 constructs a sub-model library based on operating condition clustering, and switches or fine-tunes the optimal sub-model online according to the current operating status and prediction error.

[0169] In this embodiment, the construction of the sub-model library based on working condition clustering specifically refers to:

[0170] First, clustering algorithms such as K-Means or Gaussian Mixture Model (GMM) are used to cluster the feature vectors of operating conditions in historical operating data, and similar operating conditions are grouped into the same cluster.

[0171] Then, a dedicated directional ConvLSTM sub-model is trained separately for each load case cluster, forming a sub-model library. During online runtime, the maximum posterior probability is calculated. , choose to The sub-model with the highest probability To make a prediction, the following logical representation can be used:

[0172]

[0173] In the formula, Indicates proportional to, Indicates the current input Next choice The posterior probability of each sub-model This indicates the current operating condition characteristics and the j-th cluster center. The distance between them This represents the center vector of the j-th working condition cluster. An adjustable weight is used to balance the similarity of operating conditions and the accuracy of recent predictions; Let be the root mean square error of the j-th sub-model's predictions on recent historical data.

[0174] In this embodiment, the prediction error is expressed as the root mean square error (RMSE). After the model is deployed, the system continuously monitors the RMSE of the output temperature prediction data through a rolling window. When the RMSE is within the preset warning range and the system detects a gradual change in operating conditions, the system will adjust the learning rate to a minimum. ( The system uses newly collected labeled data to fine-tune the model online to adapt to slow operating condition drift; if the RMSE suddenly increases beyond the preset warning threshold, the system's adaptive mechanism is triggered to switch the prediction model from the sub-model library.

[0175] Furthermore, if the RMSE rises above the warning threshold by more than 0.2 within 1 to 3 cycles after switching prediction models, the system will roll back to the previous sub-model and label samples for offline retraining.

[0176] S4. Based on the output of the prediction model, perform multi-objective model predictive control (MPC), construct a multi-objective cost function, and optimize the control quantity of the actuator online; wherein, the multi-objective cost function comprehensively considers temperature field tracking error, boundary gradient suppression, reaction efficiency loss, control quantity change and control quantity amplitude.

[0177] In this embodiment, the MPC internal prediction model uses the future Np-step temperature field output by S3 to predict... The system predicts NOx concentration at the outlet and ammonia emission, and combines the results from the sub-model library in real time. By optimizing the control quantities of the actuators, it optimizes the ammonia injection rate, the valve opening, and the angle of the guide baffle online.

[0178] Within each control cycle, MPC solves for a multi-objective cost function that aims to minimize the following The optimization problem is used to comprehensively achieve optimization objectives such as temperature field tracking error, boundary gradient suppression, reaction efficiency loss, control variable change, and control variable amplitude. The multi-objective cost function is expressed by the following logic. :

[0179]

[0180] In the formula, Indicates the temperature field tracking term. , , , , These are the weight coefficients for each optimization objective. Let represent the predicted temperature field at time τ. For reference or target temperature field; This represents the boundary gradient penalty term; Indicates performance loss. Represents the performance loss function; This indicates the number of prediction steps for the temperature field. Indicates the control step size. This indicates the control of incremental penalty terms. Indicates control increment; This indicates the penalty term for the control amplitude. This indicates the control amplitude.

[0181] In this embodiment, All parameters can be tuned online via the controller, with temperature and efficiency related weights. The initial range can be set to Weights related to the rate of change and magnitude of the control quantity The initial range can be set to Then, perform online adaptive updates.

[0182] This embodiment introduces a differentially efficient energy loss function that is directly related to the denitrification reaction efficiency. By incorporating the near-wall temperature difference and temperature gradient into a differentiable objective as a proxy for the denitrification reaction efficiency, the control constraints on key regions are improved. The temperature field efficiency loss function is expressed using the following logic. :

[0183]

[0184] In the formula, This means taking the average over all i grid points in the space. For the first Regional weights of spatial grid points This represents the temperature gradient at that point. The optimal reaction temperature for SCR can be obtained directly through unit design or performance testing. This is the regularization coefficient, used to penalize excessively large local temperature gradients. In this embodiment, and The efficiency of denitrification can be determined by analyzing historical data and fitting the relationship curve between temperature distribution uniformity and boundary temperature gradient, or by calibrating through segmented loading tests under different loads.

[0185] In this embodiment, since fixed weight values ​​cannot adapt to all wide-load conditions, this embodiment introduces an online adaptive weight update law, which is represented by the following logic:

[0186]

[0187] In the formula, for Spatial region weights at time points, For the clipping function, through The function limits the updated weights to a preset safety range. Inside, These represent the minimum and maximum values ​​of the spatial region weights, respectively. This is a forgetting factor, used to enable weights to slowly return to their initial values ​​and quickly adapt to new working conditions; The learning rate is used to control the speed of weight adjustment. Is with the first Performance metrics relevant to each region, such as the deviation between the predicted temperature and the optimal reaction temperature in that region. This can be used to reflect local control performance, such as the region's contribution to the overall denitrification efficiency.

[0188] Furthermore, multi-objective cost function The optimization solution also needs to satisfy constraints, including hard constraints and soft constraints; the hard constraints specifically mean that the magnitude and increment of the control quantity are strictly limited within the physically permissible range, which can be represented by the following logic:

[0189]

[0190]

[0191] In the formula, , These represent the maximum and minimum permissible values ​​of the control amplitude, respectively. , These represent the maximum and minimum allowable values ​​for the control quantity increment, respectively.

[0192] The soft constraints specifically involve imposing constraints on key output variables such as predicted ammonia slip and NOx emissions. These constraints are typically used as the objective function. This is achieved using high-weight penalty terms to prioritize environmental protection indicators while ensuring that the optimization problem always has a feasible solution. This can be represented by the following logic:

[0193]

[0194]

[0195] In the formula, This represents the predicted ammonia release at time t+τ. This indicates the maximum permissible amount of ammonia release. This represents the predicted NOx concentration at the export point at time t+τ. The emission limit for NOx concentration at the export site.

[0196] Furthermore, to meet the real-time requirements of industrial sites, this embodiment provides a preferred system configuration, specifically: timing settings, including control cycles. Rolling prediction step size Control step size A high-efficiency embedded optimization solver is selected, such as OSQP or qpOASES, to ensure that... Complete one full "prediction-optimization-execution" cycle within the time limit (that is, repeatedly execute steps S3 to S4).

[0197] In this embodiment, a multi-objective cost function is constructed with the optimization objectives of minimizing temperature field tracking error, boundary gradient suppression, reaction efficiency loss, control quantity change, and control quantity amplitude, and an economic penalty term is introduced to achieve a comprehensive optimization effect. Since catalytic reaction efficiency is a difficult-to-measure indicator, a differentiable efficiency loss function directly related to denitrification reaction efficiency is introduced. The catalytic reaction efficiency is also integrated into the optimization framework.

[0198] This embodiment also introduces an online adaptive update law for weights, enabling the weights of the MPC controller to be adaptively adjusted to adapt to wide load conditions. Furthermore, by setting multiple constraints, the control amplitude, rate of change, emission environmental indicators, and economic efficiency of the actuator are constrained, so that the control commands output to the actuator can achieve temperature field self-equilibrium, ensure denitrification efficiency, and meet economic and environmental indicators.

[0199] S5 performs actuator delay compensation on the optimized control quantity and issues it, sets rollback conditions, and automatically rolls back to the original control strategy and continues to collect data when the rollback conditions are met. Specifically, S5 includes the following:

[0200] In this embodiment, the optimized control quantity output by MPC is sent to the actuator control loop of the DCS via the OPC UA / Modbus communication protocol. The optimized control quantity is output based on Multi-Objective Model Predictive Control (MPC). The system sends data to the actuator control loop of the DCS via the OPC UA / Modbus communication protocol. The communication uses a heartbeat mechanism, sending a heartbeat packet every 100ms to monitor the communication link status between the system and the DCS in real time.

[0201] Since the control input needs to consider the nonlinear characteristics of the actuator, such as rate limitation, dead zone, and saturation, a Smith predictor is used to compensate for the total delay. This can be represented using the following logic:

[0202]

[0203] In the formula, To compensate for the delayed control commands, for The amount of control at any given moment The original control quantity calculated by MPC at the current moment. Indicates model parameters, This represents the rate of change of the historical control quantity.

[0204] In this embodiment, the rollback conditions include sliding window error warning, optimization infeasibility, communication anomaly, and interlocking trigger. The sliding window error warning specifically involves calculating the root mean square error (RMSE) between the model-predicted temperature and the measured temperature within the most recent time window (e.g., 5 minutes), which is the sliding window error. Specifically, a preset RMSE of 3 is used as the out-of-bounds threshold, and the warning threshold is the 80% out-of-bounds threshold, i.e., RMSE = 2.4. Reduced-order control can be implemented within an average of 1-3 cycles. This reduced-order control is a safe rollback strategy, meaning that when the intelligent control effect is poor, the system automatically switches to a simpler, more robust control mode. For example, switching back to the original PID control loop of the DCS, switching to manual operation mode, or switching to a conservation logic mode based on simple functions such as load and flue gas volume to calculate the ammonia injection quantity, to ensure the basic stability and safe operation of the system.

[0205] The infeasibility of optimization specifically refers to the fact that the solver in S4 cannot find the optimal solution under the given constraints, indicating that the control objectives are conflicting or the operating conditions are extremely abnormal.

[0206] The specific communication anomaly is defined as: if the heartbeat packet loss with DCS exceeds 3 consecutive control cycles (750ms), it is determined to be a communication anomaly.

[0207] The interlocking trigger specifically refers to receiving hard-wired interlocking signals such as emergency stop signals from the DCS.

[0208] In this embodiment, the automatic fallback to the original control strategy specifically means that the system stops issuing optimized control quantities based on MPC and switches back to the original PID control loop of the DCS, switches to manual operation mode, or switches to the ammonia injection conservation logic mode; wherein, the ammonia injection conservation logic mode specifically calculates and fixes the ammonia injection quantity using a simple function that is in a fixed proportion to the load or flue gas volume.

[0209] After the abnormal situation is resolved, the system will not automatically switch back to MPC optimized control. It must wait for the operating conditions to stabilize and meet the switching back conditions before it can switch back. The specific switching back conditions are: the current operating conditions have been running stably for more than 60 seconds, and all performance indicators (such as RMSE) have recovered to below the safety threshold. The safety threshold is set to the 60% out-of-bounds threshold, that is, the safety threshold is RMSE=1.8.

[0210] This invention can be applied to SCR reactors in coal-fired, biomass-fired, and co-fired power units. It is compatible with OPC UA and Modbus communication and can achieve temperature field equilibrium and denitrification / economic synergistic optimization without modifying the original DCS main circuit. It has significant energy-saving, emission-reduction, and retrofitting value.

[0211] Example 2

[0212] To verify the performance of the SCR reactor temperature field self-balancing control method proposed in this invention under steady-state conditions, this embodiment takes a 300 MW subcritical coal-fired unit under a single-condition steady-state period as an example. The SCR reactor temperature field self-balancing control method disclosed in Embodiment 1 is adopted, based on the following specific configuration:

[0213] Hardware configuration: An industrial-grade edge computing device is used as the core processing unit. The edge computing device is equipped with a CPU (8 cores, 16 threads, clock speed ≥ 2.5GHz), a GPU (TDP ≤ 70W), and 16GB of RAM, and runs on the RT-Linux real-time operating system. It employs a dual-network redundant architecture to connect to the DCS system, with one network using industrial Ethernet and the other using the OPC UA protocol. The heartbeat detection cycle is set to 100ms. Number of on-site temperature measurement points. The system consists of 32 units, deployed using a combination of thermocouples and optical fibers, and equipped with one online NOx analyzer and one NH3 analyzer.

[0214] Model configuration: The ratio of the relevant lengths in the horizontal and vertical directions in S22 is... Virtual wall point spacing m, boundary zone thickness m; In S32, the directional ConvLSTM unit has a main / laterally separable convolution kernel of 3×1 or 1×3; In S31, the historical frame K is 8; In S4, the prediction stride Np=30, the control stride Nc=8, and the sampling period Ts=250ms.

[0215] This embodiment also includes a baseline control system, employing isotropic Kriging for temperature field reconstruction and an MPC controller that considers only the temperature deviation term as a single objective. All other constraints remain consistent with the system of this invention. The system ran continuously for 7 days. The first day was used for initial parameter calibration and system warm-up; data from this day was not included in the statistics. The following 6 days were the formal testing period, during which all operating data were recorded, and the compliance range remained stable at 70% ± 5%.

[0216] This embodiment evaluates the performance of the output results and the baseline comparison system using the following metrics: prediction accuracy, balance, economic / environmental indicators, real-time performance, and stability. Prediction accuracy includes the root mean square error (RMSE_T) of the temperature field in the SCR inlet region and the root mean square error (RMSE_edge) of the near-wall region. Balance includes the temperature field variance (Var_T) and the 95th quantile (P95) of the near-wall gradient. Economic / environmental indicators include daily ammonia injection consumption and ammonia slip proxy (A_slip). Real-time performance and stability include end-to-end delay decomposition and MPC infeasibility rate.

[0217] Compared to the baseline control system, the root mean square error (RMSE) of the temperature field in this embodiment can be reduced by approximately 15% to 30%, the root mean square error (RMSE) edge of the near-wall region can be reduced by approximately 20% to 40%, the variance of the temperature field (Var) can be reduced by approximately 15% to 25%, the 95th percentile of the wall gradient (P95) can be reduced by approximately 10% to 20%, the average daily ammonia injection consumption can be reduced by approximately 5% to 12%, the ammonia slip proxy (A_slip) can be reduced by approximately 10% to 20%, the end-to-end delay (95) can be ≤350ms, the communication delay (95) can be ≤200ms, the average QP solution time can be ≤35ms, and the MPC infeasibility rate can be less than 0.5‰.

[0218] Example 3

[0219] To verify the dynamic response capability and the effectiveness of the online switching mechanism of the sub-model library under conditions of large and rapid changes in unit load (deep peak shaving condition), this embodiment uses a 600MW supercritical coal-fired unit as the test object and adopts the SCR reactor temperature field self-balancing control method disclosed in Embodiment 1. The test condition is set to simulate a deep peak shaving command from the power grid, and the following load change test is performed:

[0220] Rapid load reduction phase: Within 30 minutes, the unit's load rapidly decreases from 90% (approximately 540MW) to 40% (approximately 240MW).

[0221] Low-load stabilization phase: Stable operation for 1 hour at 40% load.

[0222] Rapid load ramp-up phase: Within 40 minutes, the load climbs from 40% back to 80% (approximately 480MW).

[0223] Sub-model library configuration: The sub-model library (step S34) in this embodiment is pre-trained offline and stores three sub-models, including a high-load sub-model (suitable for working conditions with load > 65%), a medium-low load sub-model (suitable for working conditions with load ≤ 65%), and a variable load sub-model (suitable for high dynamic processes with load change rate > 2% / min).

[0224] The baseline control system used for comparison is a traditional ConvLSTM combined with MPC system using a set of global parameters, without a condition switching mechanism.

[0225] Analysis of sub-model switching behavior: During the load reduction phase, when the load crosses the approximately 65% ​​load point, the system of this invention can automatically and smoothly switch from the "high load sub-model" to the "medium-low load sub-model" based on the real-time operating conditions, and the entire switching process is free of control disturbances; During the load increase phase, when the load crosses the approximately 55% load point (considering hysteresis effects or different switching thresholds), the system smoothly switches back to the "high load sub-model".

[0226] Dynamic performance comparison analysis: During the load reduction phase, the baseline system exhibited severe prediction bias due to model mismatch, leading to excessive NOx outlet concentration (peak value exceeding the limit by 15%) for approximately 8 minutes. Furthermore, to ensure a basic denitrification rate, the instantaneous peak ammonia slip reached 8 ppm. In contrast, the system in this embodiment, through timely model switching and proactive MPC adjustment, maintained stable NOx concentrations within the limits, with ammonia slip peaks not exceeding 4.5 ppm. During the low-load stabilization phase, the baseline system showed deterioration in control quality under low load, resulting in a large temperature field standard deviation. The system in this embodiment maintained a temperature field standard deviation approximately 3°C lower than the baseline system, closer to the center of the catalyst activity temperature window, effectively suppressing the risk of ammonium bisulfate formation. During the load increase phase, the system in this embodiment reached a new steady state (80% load) approximately 12 minutes faster than the baseline system, with a lower control overshoot.

[0227] As can be seen from the above, this invention, based on a clustering sub-model library for operating conditions and an online switching mechanism, can effectively cope with the strong nonlinear dynamics caused by drastic load changes under deep peak-shaving conditions. Compared with a single global model, this invention can achieve faster and smoother control transitions, significantly improving the stability and compliance of emission control during dynamic processes, demonstrating outstanding industrial application value.

[0228] This invention has been verified on coal-fired power units of different capacities and operating conditions. Its system architecture is compatible with industrial standard communication protocols such as OPC UA and Modbus. It can be integrated with existing DCS in an external design. Without changing the self-balancing control logic of the SCR reactor temperature field, it can realize the safe, efficient and economical intelligent operation of the SCR system. It has broad prospects for promotion and significant industrial practical value.

[0229] Example 4

[0230] This invention also provides a diagnostic system for the above-mentioned self-equilibrium control method of temperature field in an SCR reactor based on directional ConvLSTM, comprising:

[0231] The data processing module is used to collect sparse temperature measurement data from the SCR reactor and perform data preprocessing.

[0232] The interpolation and reconstruction module is used to perform spatial interpolation and reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features.

[0233] The model prediction module combines the output of spatial interpolation reconstruction with operating condition data, coal type information, and actuator status data as input features to construct a directional ConvLSTM-based spatiotemporal prediction model. The model is trained using a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration, and ammonia escape amount for multiple future steps.

[0234] The MPC module performs multi-objective model predictive control based on the output of the predictive model, constructs a multi-objective cost function, and optimizes the control quantity of the actuator online.

[0235] The control execution module is used to compensate for the delay of the optimized control quantity and issue it, set the rollback conditions, and automatically roll back to the original control strategy when the rollback conditions are met and continue to collect data.

[0236] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM, characterized in that, Includes the following steps: S1, Collect sparse temperature measurement data of the denitrification reactor and perform data preprocessing; S2 performs spatial interpolation reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features. S3, based on the output of spatial interpolation reconstruction, combined with operating condition data, coal type information, and actuator status data as input features, constructs a spatiotemporal prediction model based on directional ConvLSTM. The model is trained using a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration, and ammonia escape amount for future multiple steps. The model includes a shared encoder and a multi-task decoder. The shared encoder is composed of L layers of stacked directional ConvLSTM units, with directional weights introduced during the update process of each directional ConvLSTM unit. , : In the formula, Let be the hidden state at time t. and are one-dimensional convolution operations along the x and y directions, respectively, and b is the bias term; S4, based on the output of the prediction model, performs multi-objective model predictive control, constructs a multi-objective cost function, and optimizes the control quantity of the actuator online; S5 performs actuator delay compensation on the optimized control quantity and issues it, sets rollback conditions, and automatically rolls back to the original control strategy when the rollback conditions are met and continues to collect data.

2. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 1, characterized in that, S1 includes the following steps: S11, collect sparse temperature data of the denitrification reactor, including temperature, valve position, load, total flue gas volume, outlet NOx concentration, outlet NH3 concentration and outlet SO2 concentration; S12 performs data preprocessing on the collected sparse temperature measurement data, including data time alignment, data cleaning, and data normalization.

3. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 1, characterized in that, S2 includes the following steps: S21, Define the wall grid and calibrate the coordinates of the measuring points; S22, spatial interpolation reconstruction of sparse temperature measurement data based on the anisotropic Kriging interpolation algorithm, specifically: Constructing anisotropic covariance functions As a kernel function, it is represented using the following logical expression: In the formula, Let be the spatial displacement vector between two points. Indicates to Transpose operation; The signal variance characterizes the overall fluctuation amplitude of the temperature field; To measure the noise variance, Indicates exponentiation. For the Dirac function, An anisotropic metric tensor, used to define the shape and orientation of spatial correlation, is represented by the following logic. : In the formula, The relevant lengths along the principal axis, It is a rotation matrix, dominated by the flue gas direction angle. This is determined to align the principal axis of the covariance ellipsoid with the direction of flue gas flow. Kriging estimation can be expressed using the following logic: In the formula, Let x be the temperature value to be predicted at location point x. A column vector consisting of temperature observations from all N actual temperature measurement points. , Given the covariance matrix between the measurement points, Points to be interpolated The covariance vector between the known measurement points and the total covariance vector. S23, Set virtual wall points and apply boundary gradient constraints; The specific steps of setting virtual wall points are as follows: placing virtual wall points at preset intervals along the wall surface, and setting the virtual wall surface temperature. And the uncertainty of virtual wall temperature ;in, This represents the variance of the actual measured temperature values ​​in the near-wall region. for The corresponding weighting coefficients; The boundary gradient constraint specifically refers to: reconstructing the gradient of the temperature field along the wall normal direction. No more than one maximum allowed value This can be represented using the following logic: In the formula, Represents the temperature field. Indicates the direction of the wall normal. Indicates the boundary gradient constraint value. This indicates the preset heat transfer limit value; S24, update the hyperparameters of the Kriging interpolation algorithm online using maximum likelihood estimation. Record hyperparameters The log-likelihood function is obtained by taking the logarithm of the maximum likelihood estimation function. The log-likelihood function is then expressed using the following logic. : In the formula, Representing the covariance matrix, exponential smoothing filtering is used to smooth the hyperparameters. Periodic online updates are performed using the following logical representation: In the formula, These are the hyperparameters updated at time t. This is the maximum likelihood estimate at the current moment. This is the smoothing coefficient.

4. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 1, characterized in that, S3 includes the following steps: S31, Define the control vector Coal type indicator vector and working condition vector The input tensor of the prediction model is constructed by combining the output of step S2. ;in, , , The output of step S2 consists of a three-channel image sequence comprising the temperature field, boundary gradient field, and near-wall mask of the past K frames; the input tensor is... Total number of input channels Recorded as Where K represents the number of historical frames, Indicates the vector dimension; S32, Construct a spatiotemporal prediction model based on directional ConvLSTM; The multi-task decoder includes a temperature prediction head and an emission prediction head. The temperature prediction head includes several deconvolutional layers or upsampling layers to decode the encoded spatiotemporal features into a temperature field sequence for the next few steps. The emission prediction head includes a global pooling layer and a fully connected layer to regress and predict the future NOx concentration and NH3 escape. S33, during the model training phase, uses a boundary-first masking strategy to handle missing tests and fault points, specifically by using a masking strategy with a higher probability than the core region. Higher probability Randomly select the near-wall region of the input tensor. The block is masked by setting all feature values ​​within the block to 0, forcing the prediction model to rely on contextual information and boundary features to infer the value of the missing point; S34, construct a sub-model library based on operating condition clustering, and switch or fine-tune the optimal sub-model online according to the current operating status and prediction error; the construction of the sub-model library based on operating condition clustering specifically involves: First, clustering algorithms are used to cluster the feature vectors of operating conditions in historical operating data, and similar operating conditions are grouped into the same cluster. Then, a dedicated directional ConvLSTM sub-model is trained separately for each load case cluster, forming a sub-model library. During online runtime, the maximum posterior probability is calculated. , choose to The sub-model with the highest probability To make a prediction, the following logical representation can be used: In the formula, Indicates proportional to, Indicates the current input Next choice The posterior probability of each sub-model This indicates the current operating condition characteristics and the j-th cluster center. The distance between them This represents the center vector of the j-th working condition cluster. Adjustable weights are used to balance the similarity of operating conditions and the accuracy of recent predictions; Let be the root mean square error of the j-th sub-model's predictions on recent historical data.

5. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 4, characterized in that, The direction weight in S32 , By setting up a small gating network Dynamically generated based on the current operating conditions, using the following logical representation: In the formula, For the Sigmoid function, The dominant direction of flue gas.

6. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 4, characterized in that, The prediction model in S32 employs a multi-task loss function based on uncertainty, which is represented by the following logic: : In the formula, The predicted losses are for temperature, NOx, and NH3, respectively. These are the uncertainty terms for temperature, NOx, and NH3, respectively.

7. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 1, characterized in that, The multi-objective cost function is represented in S4 using the following logic. : In the formula, Indicates the temperature field tracking term. , , , , These are the weight coefficients for each optimization objective. Let represent the predicted temperature field at time τ. For reference or target temperature field; This represents the boundary gradient penalty term; Indicates performance loss. Represents the performance loss function; This indicates the number of prediction steps for the temperature field. Indicates the control step size. This indicates the control of incremental penalty terms. Indicates control increment; This indicates the penalty term for the control amplitude. The control amplitude is represented; the temperature field efficiency loss function is represented using the following logic. : In the formula, This means taking the average over all i grid points in the space. For the first Regional weights of spatial grid points This represents the temperature gradient at that point. The optimal reaction temperature for SCR is... is the regularization coefficient.

8. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 7, characterized in that, The Adjustments are made by introducing an online adaptive update law for the weights, using the following logical representation: In the formula, for Spatial region weights at time points, For the clipping function, through The function limits the updated weights to a preset safety range. Inside, These represent the minimum and maximum values ​​of the spatial region weights, respectively. Forgetting factor, For learning rate, It is the first Several performance metrics.

9. The self-equilibrium control method for the temperature field of a denitrification reactor based on directional ConvLSTM according to claim 7, characterized in that, The multi-objective cost function in S4 The optimization solution applies constraints, including hard constraints and soft constraints; the hard constraints specifically mean that the magnitude and increment of the control quantity are strictly limited to the physically permissible range, which is represented by the following logic: In the formula, , These represent the maximum and minimum permissible values ​​of the control amplitude, respectively. , These represent the maximum and minimum allowable values ​​for the control quantity increment, respectively; The soft constraints specifically involve constraining key output variables such as predicted ammonia slip and NOx emissions, and are represented by the following logic: In the formula, This represents the predicted ammonia release at time t+τ. This indicates the maximum permissible amount of ammonia release. This represents the predicted NOx concentration at the export point at time t+τ. The emission limit for NOx concentration at the export site.

10. A self-balancing control system for the temperature field of a denitrification reactor based on directional ConvLSTM, characterized in that, include: The data processing module is used to collect sparse temperature measurement data from the denitrification reactor and perform data preprocessing. The interpolation and reconstruction module is used to perform spatial interpolation and reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features. The model prediction module, based on the output of spatial interpolation reconstruction combined with operating condition data, coal type information, and actuator status data as input features, constructs a spatiotemporal prediction model based on directional ConvLSTM. The model is trained using a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration, and ammonia escape for multiple future steps. The model includes a shared encoder and a multi-task decoder. The shared encoder consists of L layers of stacked directional ConvLSTM units, with directional weights introduced during the update process of each directional ConvLSTM unit. , : In the formula, Let be the hidden state at time t. and are one-dimensional convolution operations along the x and y directions, respectively, and b is the bias term; The MPC module performs multi-objective model predictive control based on the output of the predictive model, constructs a multi-objective cost function, and optimizes the control quantity of the actuator online. The control execution module is used to compensate for the delay of the optimized control quantity and issue it, set the rollback conditions, and automatically roll back to the original control strategy when the rollback conditions are met and continue to collect data.

Images