Ship exhaust gas denitration intelligent control system and method thereof

Abstract

The application provides a ship tail gas denitration intelligent control system and method, relates to the technical field of intelligent control, and solves the problems of poor variable condition adaptability and difficult inhibition of ammonia escape in the prior art. The method comprises the following steps: constructing a gray-box engine source nitrogen oxide predictor, predicting the nitrogen oxide concentration at the reactor inlet according to real-time operating parameters, and generating feedforward information; constructing a virtual ammonia coverage observer based on extended Kalman filtering, and estimating the ammonia coverage on the catalyst surface in real time by using sensor signals; constructing a model predictive controller based on ammonia coverage constraints, taking the feedforward information as a reference, taking the ammonia coverage as a state feedback, and optimizing the urea injection instruction online to inhibit ammonia escape; and finally, injecting urea according to the instruction to realize catalytic reduction denitration.

Description

Technical Field

[0001] This invention belongs to the field of ship exhaust gas treatment and intelligent control technology, specifically a ship exhaust gas denitrification intelligent control system and method. Background Technology

[0002] Nitrogen oxides (NOx) in ship exhaust are one of the key pollutants controlled by the International Maritime Organization (IMO). Selective catalytic reduction (SCR) technology has become the mainstream solution to meet IMO Tier III emission standards due to its high denitrification efficiency and mature technology.

[0003] Existing SCR control methods typically employ a PID control strategy based on NOx sensor feedback. This involves calculating the urea injection rate using a proportional-integral-derivative controller based on NOx sensor measurements downstream of the SCR reactor. However, this method has significant drawbacks: First, the NOx sensor itself exhibits measurement lag (typically several to tens of seconds), causing the control response to lag behind actual operating conditions. This can lead to excessive or insufficient ammonia injection during frequent load changes on ships. Second, existing methods cannot directly measure the ammonia coverage on the catalyst surface, a crucial state variable determining denitrification efficiency and ammonia escape risk. This lack of awareness of the internal state makes control decisions somewhat arbitrary. Third, traditional control strategies do not adequately consider ammonia escape constraints. During sudden changes in operating conditions, excessive ammonia injection can easily lead to excessive ammonia escape peaks, wasting reducing agent and causing secondary pollution. Furthermore, ship engines operate at low exhaust temperatures under low load conditions, resulting in a significant decrease in SCR catalyst activity. Existing methods lack effective countermeasures for low-temperature conditions, leading to a sharp drop in denitrification efficiency. Therefore, how to achieve rapid prediction of source NOx concentration, accurate estimation of ammonia coverage on catalyst surface, and optimal ammonia injection control under constraints are technical problems that urgently need to be solved in this field. Summary of the Invention

[0004] This application provides an intelligent control system and method for denitrification of ship exhaust gas, which solves the technical problems of existing technologies such as slow response, unmeasurable ammonia coverage, poor adaptability to changing operating conditions, difficulty in suppressing ammonia escape, and low denitrification efficiency at low temperatures.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a smart control method for denitrification of ship exhaust gas is provided, applied to a ship exhaust gas aftertreatment system, wherein the aftertreatment system includes a selective catalytic reduction reactor and a urea injection device, and the specific method is as follows: An engine source nitrogen oxide predictor based on a gray box model is constructed and executed to predict the source nitrogen oxide concentration before entering the selective catalytic reduction reactor based on the acquired real-time engine operating parameters, and to generate feedforward information. A virtual ammonia coverage observer based on extended Kalman filtering is constructed and executed to estimate the ammonia coverage on the catalyst surface in the selective catalytic reduction reactor in real time based on the inlet and outlet sensor signals of the selective catalytic reduction reactor. A model predictive controller based on ammonia coverage constraint is constructed and executed. The feedforward information is used as the reference input, the ammonia coverage on the catalyst surface is used as the state feedback, and the urea injection command is obtained through online rolling optimization calculation with the constraint objective of suppressing ammonia escape. Based on the urea injection command, the urea injection device is controlled to inject urea into the exhaust gas to catalytically reduce nitrogen oxides in the exhaust gas.

[0006] Based on the above technical solutions, this application provides an intelligent control method for ship exhaust gas denitrification. By constructing and executing an engine source nitrogen oxide predictor based on a gray box model, it achieves accurate feedforward prediction of the source nitrogen oxide concentration before entering the selective catalytic reduction reactor, overcoming the response delay problem caused by sensor measurement lag in traditional control methods. By constructing and executing a virtual ammonia coverage observer based on extended Kalman filtering, it uses existing sensor signals to estimate the intangible ammonia coverage on the catalyst surface in real time, obtaining key state information without adding additional hardware. The model predictive controller based on ammonia coverage constraints combines feedforward information with state feedback and performs online rolling optimization with the constraint objective of suppressing ammonia escape. This achieves accurate calculation of urea injection commands, effectively suppressing the risk of ammonia escape while ensuring high denitrification efficiency. The coordinated cooperation of the predictor, observer, and controller forms an integrated intelligent control architecture of feedforward-feedback-optimization, which significantly improves the response speed, control accuracy, and operational stability of the ship's exhaust gas denitrification system under varying operating conditions, reduces the risk of ammonia escape and catalyst poisoning, and has outstanding advantages such as precise control, low hardware dependence, and strong adaptability.

[0007] Furthermore, the selective catalytic reduction reactor is a device for providing a catalytic reaction site, used to enable nitrogen oxides and reducing agents to undergo a selective catalytic reduction reaction under the action of a catalyst to generate nitrogen gas and water; The real-time operating parameters include engine speed, engine load, fuel injection quantity, boost pressure, exhaust temperature, exhaust flow rate, and excess air coefficient; wherein, the excess air coefficient represents the ratio of the actual mass of air entering the internal combustion engine cylinder to the mass of air required for theoretical complete combustion.

[0008] Furthermore, the intelligent control method for ship exhaust gas denitrification also includes acquiring predicted navigation condition information and pre-adjusting the active thermal state of the selective catalytic reduction reactor based on the navigation condition information, specifically including: A thermal inertial model of a selective catalytic reduction reactor is constructed. The thermal inertial model describes the dynamic response characteristics of the catalyst bed temperature as a function of exhaust temperature and active heating power using first-order inertia plus a pure delay element. Based on the predicted navigation condition information and the thermal inertia model, the catalyst bed temperature of the selective catalytic reduction reactor is predicted within a preset time period in the future, and the predicted temperature trajectory is obtained. If there is a range in the predicted temperature trajectory that is lower than the preset low temperature threshold, the heating device connected to the selective catalytic reduction reactor will be started in advance at the current moment so that the catalyst bed reaches the preset target temperature range before entering the low temperature condition.

[0009] Furthermore, the method for obtaining the predicted navigation condition information is as follows: the ship's speed change trajectory, port entry and exit plan, or engine load change trajectory within a preset time period are obtained through the ship navigation planning system or the ship autopilot system, and used as the predicted navigation condition information.

[0010] Furthermore, the engine source nitrogen oxide predictor adopts a gray box modeling architecture, including a mechanism model part and a black box correction model part; The mechanistic model is based on the extended Zeldovich mechanism. It uses the combustion characteristic parameters in the engine's real-time operating parameters to calculate the in-cylinder characteristic temperature and calculates the nitrogen oxide generation rate based on the Arrhenius equation to obtain the mechanistic model output value. The black-box correction model part adopts a radial basis function network. Taking the real-time operating parameters of the engine as input, it outputs the deviation correction value between the output value of the mechanism model and the actual source nitrogen oxide concentration, and obtains the output value of the black-box correction model, which is used to compensate for the prediction deviation caused by combustion non-ideal in the mechanism model part. The engine source nitrogen oxide predictor obtains the predicted source nitrogen oxide concentration by adding the output value of the mechanism model to the output value of the black box correction model.

[0011] Furthermore, the parameters of the black-box correction model are obtained through offline identification. The offline identification aims to minimize the root mean square error between the predicted value and the measured value, and introduces a regularization term to prevent overfitting and ensure the model's extrapolation ability outside the training data domain. The regularization term includes a weight decay term and a gradient constraint term. The gradient constraint term is used to force the gradient of the black-box correction model part with respect to the engine operating parameters to remain orthogonal to the gradient of the mechanism model part with respect to the engine operating parameters, so that the black-box correction model part only compensates for the nonlinear residuals not captured by the mechanism model part, without changing the main physical law of nitrogen oxide generation changing with operating conditions reflected by the mechanism model part.

[0012] Furthermore, the objective function expression for the offline identification is: ;in, This represents the measured concentration of source nitrogen oxides at the k-th sampling point. This is the engine operating parameter vector for the k-th sampling point. This is the mechanistic model part based on the extended Zeldovich mechanism. This is the part of the black-box correction model based on radial basis function networks. The parameters to be identified include the center vector, width parameter, and output layer weights of the radial basis function network; This is the weight decay regularization coefficient. These are the gradient constraint regularization coefficients. To correct the gradient of the black-box model with respect to engine operating parameters, This represents the gradient of the mechanistic model with respect to engine operating parameters; This is used to force the gradient of the black-box correction model to remain orthogonal to the gradient of the mechanistic model.

[0013] Furthermore, the virtual ammonia coverage observer is built on an extended Kalman filter framework and includes a state prediction module, an observation update module, and a sensor signal decoupling module. The virtual ammonia coverage observer establishes a discrete state space model. The discrete state space model uses the ammonia coverage on the catalyst surface and the gas phase ammonia concentration in the selective catalytic reduction reactor as state variables, the ammonia concentration corresponding to the urea injection rate as the control input, and the exhaust temperature and exhaust flow rate as time-varying parameters to construct a state transition equation. The sensor signal decoupling module constructs an observation equation based on the cross-sensitivity characteristics of the nitrogen oxide sensor at the outlet of the selective catalytic reduction reactor to ammonia. The observation equation adopts a dual-timescale dynamic separation model, which expresses the measurement signal of the nitrogen oxide sensor as the superposition of the actual nitrogen oxide concentration response and the actual ammonia concentration response at the outlet. The actual nitrogen oxide concentration response is characterized by a first time constant, and the actual ammonia concentration response is characterized by a second time constant, and the second time constant is smaller than the first time constant. The state prediction module and the observation update module perform extended Kalman filter recursion based on the state transition equation and the observation equation, output the estimated value of ammonia coverage on the catalyst surface in real time, and output the estimated value as a state feedback signal to the model prediction controller.

[0014] Furthermore, the observation equation includes a continuous-time domain expression and a discrete-time domain expression; The continuous-time domain expression is: ;in, The measurement signal is from the nitrogen oxide sensor. This represents the actual nitrogen oxide concentration. The actual ammonia concentration is given, and α is the cross-sensitivity coefficient. The sensor's response time constant to nitrogen oxides is... Let be the sensor's response time constant to ammonia, and , To measure noise, * indicates convolution operation, and t indicates time variable; The discrete-time domain expression is: ;in, , , The sampling period is This is a recursive term for the ammonia response at the previous moment. This is the measurement signal from the nitrogen oxide sensor at the current moment. This is the measurement signal from the nitrogen oxide sensor at the previous moment. This represents the current actual nitrogen oxide concentration. The value represents the current ammonia concentration, and k represents the sampling time number.

[0015] Furthermore, the model predictive controller is built on a rolling time-domain optimization framework, including a predictive model module, a constraint processing module, and an optimization solution module; The prediction model module establishes an ammonia storage kinetic model for predictive control. The ammonia storage kinetic model describes the dynamic change of ammonia coverage on the catalyst surface using a first-order equation of state. The ammonia concentration corresponding to the urea injection command is used as the control input, the source nitrogen oxide concentration corresponding to the feedforward information is used as the known disturbance input, and the ammonia coverage on the catalyst surface is used as the current value of the state variable. In each control cycle, the constraint processing module calculates the upper limit constraint value of the ammonia coverage on the catalyst surface based on the currently measured exhaust temperature and space velocity, and sets the control input amplitude constraint. The optimization solution module takes minimizing the sum of squares of the deviation between the predicted outlet nitrogen oxide concentration and the target in the time domain as the optimization objective, and constructs and solves the rolling time domain optimization problem with the control input amplitude constraint and the upper bound constraint as the constraint conditions. The first element of the optimal control sequence obtained by the solution is used as the urea injection command output of the previous time step.

[0016] Furthermore, when the constraint processing module calculates the upper constraint value of ammonia coverage on the catalyst surface, it is based on the adsorption reaction rate constant, desorption reaction rate constant, saturated gas phase ammonia concentration, and space velocity parameter. The upper constraint value increases with increasing exhaust temperature and decreases with increasing space velocity.

[0017] Furthermore, the formula for calculating the upper limit constraint value of the ammonia coverage on the catalyst surface is as follows: ;in, This represents the maximum permissible ammonia coverage under current operating conditions. Where T is the saturated ammonia coverage and T is the exhaust temperature. Airspeed, This refers to the exhaust volume flow rate. For catalyst volume, This represents the saturated gaseous ammonia concentration. Let be the adsorption reaction rate constant. Let R be the desorption reaction rate constant, and R be the ideal gas constant. This is the pre-exponential factor for the adsorption reaction. The activation energy of the adsorption reaction. It is the pre-exponential factor for the desorption reaction. This is the activation energy for the desorption reaction; In each control cycle of the model predictive controller, based on the currently measured exhaust temperature... and airspeed Real-time computing And this is added as a hard constraint to the rolling time-domain optimization problem: ;in, To predict the ammonia coverage value at step i in the time domain, To predict the length of the time domain.

[0018] Secondly, this application provides an intelligent control system for denitrification of ship exhaust gas, comprising: an engine-source nitrogen oxide prediction module, a virtual ammonia coverage observation module, and a model predictive control module; wherein, The engine source nitrogen oxide prediction module is used to predict the source nitrogen oxide concentration before entering the selective catalytic reduction reactor based on the acquired real-time engine operating parameters, and generate feedforward information. The virtual ammonia coverage observation module is used to estimate the ammonia coverage on the catalyst surface in the selective catalytic reduction reactor in real time based on the inlet and outlet sensor signals of the selective catalytic reduction reactor. The model prediction and control module is used to take the feedforward information as a reference input, the ammonia coverage on the catalyst surface as a state feedback, and take suppressing ammonia escape as a constraint objective to obtain urea injection commands through online rolling optimization calculations, so as to control the urea injection device to inject urea into the exhaust gas.

[0019] Compared with the prior art, the beneficial effects of this application are: This application constructs an engine-source nitrogen oxide predictor based on a gray-box model, integrating a mechanistic model based on the extended Zeldovich mechanism with a black-box correction model based on radial basis function networks. This preserves the physical laws governing nitrogen oxide generation while compensating for non-ideal deviations in the actual combustion process through the black-box correction model. Furthermore, by introducing a gradient-constrained regularization term for offline identification, the black-box correction model only compensates for nonlinear residuals not captured by the mechanistic model, without altering the primary physical trends reflected by the mechanistic model. This significantly improves the extrapolation capability and generalization performance of the prediction model under different operating conditions. Simultaneously, this application fully considers the dynamic characteristics of ship navigation. By acquiring predicted navigation condition information and constructing a thermal inertia model, it achieves proactive pre-regulation of the thermal state of the selective catalytic reduction reactor. The heating device is activated in advance before the onset of low-temperature conditions, effectively avoiding catalyst deactivation due to low temperatures and ammonium bisulfate condensation and blockage, thus significantly improving the operational reliability of the aftertreatment system under varying operating conditions.

[0020] This application constructs a virtual ammonia coverage observer based on extended Kalman filtering. Utilizing the cross-sensitivity of ammonia to the nitrogen oxide sensor at the outlet of the selective catalytic reduction reactor, an observation equation is constructed using a dual-timescale dynamic separation model. The sensor measurement signal is decoupled into a superposition of the actual nitrogen oxide concentration response and the actual ammonia concentration response. Extended Kalman filtering recursion is then performed based on the state transition equation and the observation equation, achieving real-time and accurate estimation of the ammonia coverage on the catalyst surface, which cannot be directly measured. This eliminates the need for additional ammonia sensors to obtain key state information, reducing system hardware costs. Furthermore, this application constructs a model predictive controller based on rolling time-domain optimization. Feedforward information is used as the reference input, and the ammonia coverage estimate is used as the state feedback. A rolling optimization solution is performed using a dynamic upper bound of ammonia coverage as a hard constraint. This dynamic upper bound increases with increasing exhaust temperature and decreases with increasing space velocity, adaptively managing the safe ammonia storage capacity under different operating conditions. This effectively suppresses ammonia escape risk while ensuring high denitrification efficiency, achieving integrated intelligent control of feedforward-feedback-optimization.

[0021] This application establishes an intelligent control architecture through the coordinated operation of the aforementioned predictor, observer, and controller. The predictor provides feedforward information to address the sensor measurement lag problem, the observer provides key state feedback to address the unmeasurable ammonia coverage problem, and the controller makes optimal injection decisions within a constrained framework. The organic combination of these three components significantly improves the response speed, control accuracy, and operational stability of the ship's exhaust gas denitrification system under complex operating conditions, reduces the risk of catalyst poisoning and ammonia escape emissions, and has outstanding advantages such as precise control, low hardware dependence, strong adaptability to operating conditions, and high system reliability. It is particularly suitable for complex application scenarios such as navigation in emission control zones, frequent changes in operating conditions, and future ammonia fuel engines. Attached Figure Description

[0022] 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.

[0023] Figure 1 A system architecture diagram of a ship exhaust gas denitrification intelligent control system provided in this application embodiment; Figure 2 A schematic flowchart of a ship exhaust gas denitrification intelligent control method provided in an embodiment of this application; Figure 3 A schematic flowchart of another intelligent control method for ship exhaust gas denitrification provided in this application embodiment; Figure 4 A flowchart illustrating another intelligent control method for ship exhaust gas denitrification provided in this application embodiment; Figure 5 This is a flowchart illustrating another intelligent control method for ship exhaust gas denitrification provided in an embodiment of this application. Detailed Implementation

[0024] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0025] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0026] The intelligent control method for denitrification of ship exhaust gas provided in this application embodiment can be applied to, for example... Figure 1 In a ship exhaust gas denitrification intelligent control system shown, such as Figure 1As shown, the system includes: an engine-source nitrogen oxide prediction module, a virtual ammonia coverage observation module, and a model prediction and control module; among which, The engine source nitrogen oxide prediction module is used to predict the source nitrogen oxide concentration before entering the selective catalytic reduction reactor based on the acquired real-time engine operating parameters, and generate feedforward information. The virtual ammonia coverage observation module is used to estimate the ammonia coverage on the catalyst surface in the selective catalytic reduction reactor in real time based on the inlet and outlet sensor signals of the reactor. The model predictive control module uses feedforward information as reference input, ammonia coverage on the catalyst surface as state feedback, and suppresses ammonia escape as a constraint objective to obtain urea injection commands through online rolling optimization calculations, thereby controlling the urea injection device to inject urea into the exhaust gas.

[0027] To address the technical problems of poor control performance in existing ship exhaust gas denitrification systems, such as lag in response, insufficient accuracy in adapting to operating conditions, and the inability to directly measure key state variables, this application provides an intelligent control method for ship exhaust gas denitrification. This method is applied to a ship exhaust gas aftertreatment system, which includes a selective catalytic reduction reactor and a urea injection device. The specific method is as follows: A gray box model-based engine source nitrogen oxide predictor is constructed and executed to predict the source nitrogen oxide concentration before entering the selective catalytic reduction reactor based on the acquired real-time engine operating parameters, and to generate feedforward information. A virtual ammonia coverage observer based on extended Kalman filtering was constructed and executed to estimate the ammonia coverage on the catalyst surface in the selective catalytic reduction reactor in real time based on the inlet and outlet sensor signals of the reactor. A model predictive controller based on ammonia coverage constraint was constructed and executed. Feedforward information was used as reference input, ammonia coverage on catalyst surface was used as state feedback, and suppression of ammonia escape was used as the constraint objective. Urea injection command was obtained through online rolling optimization calculation. The urea injection device is controlled by a urea injection command to inject urea into the exhaust gas in order to catalytically reduce nitrogen oxides in the exhaust gas.

[0028] Based on this, this method achieves precise control of the ship's exhaust gas denitrification system under all operating conditions through a two-layer control architecture combining feedforward prediction and feedback correction. On the one hand, the source nitrogen oxide predictor based on the gray box model can detect changes in engine operating conditions in advance, overcoming the lag in response of traditional control strategies. On the other hand, the virtual observation technology based on extended Kalman filtering solves the engineering problem that ammonia coverage cannot be directly measured, while model predictive control based on ammonia coverage constraints effectively suppresses ammonia escape while ensuring high denitrification efficiency, achieving synergistic optimization of denitrification performance and environmental protection requirements.

[0029] like Figure 2 As shown in the embodiment of this application, a smart control method for denitrification of ship exhaust gas includes: S1. Construct and execute an engine source nitrogen oxide predictor based on a gray box model. Based on the obtained real-time engine operating parameters, predict the source nitrogen oxide concentration before entering the selective catalytic reduction reactor and generate feedforward information.

[0030] The gray-box model is a modeling approach that falls between the white-box and black-box models. White-box models are entirely based on physicochemical mechanisms, offering high accuracy but involving numerous parameters and complex calibration. Black-box models rely solely on fitting input-output data, neglecting internal mechanisms and exhibiting limited extrapolation capabilities. The gray-box model combines known physical mechanisms with data-driven methods, using actual operational data to identify and correct unknown parameters or unmodeled dynamics within a mechanistic model framework, thus achieving both physical interpretability and engineering practicality.

[0031] The purpose of constructing an engine source nitrogen oxide (NOx) predictor is to address the response lag problem in selective catalytic reduction (SCR) systems. Due to the gas transport delay from engine exhaust to the SCR reactor inlet, and the fact that urea injection requires hydrolysis, pyrolysis, diffusion, and adsorption processes before participating in the reaction, controlling the system solely based on the measured NOx concentration at the current moment will inevitably result in a lag. By establishing a mapping relationship between engine operating parameters such as speed, load, air-fuel ratio, and exhaust gas recirculation rate and source NOx emissions, the trend of NOx concentration changes entering the reactor can be predicted in advance.

[0032] Feedforward information refers to the basic urea injection quantity command generated in advance based on prediction results without waiting for feedback signals. Feedforward control can compensate for measurable disturbances before they affect the controlled variable, thereby significantly improving the system's response speed.

[0033] In some implementations, the engine-source nitrogen oxide (NOx) predictor can be constructed using a gray-box modeling method based on model identification. Specifically, firstly, based on the combustion thermodynamics of marine diesel engines, a simplified model framework for NOx formation is established, clarifying the theoretical relationship between key parameters such as engine speed, cyclic fuel injection quantity, intake air temperature, and intake air pressure and NOx emissions. This framework defines the basic mathematical form of the model but retains several parameters to be identified, such as the influence coefficients of various variables, time constants, and delay times. Subsequently, operating data covering different steady-state and transient operating conditions is collected during engine bench tests, including real-time parameters output by the engine's electronic control unit and measured NOx concentrations at the exhaust manifold. Using the collected data, unknown parameters in the model are identified online or offline through least squares, maximum likelihood estimation, or recursive identification algorithms, making the model output approximate actual emission values. The model after parameter identification constitutes the engine-source NOx predictor, with real-time engine operating parameters as input and predicted source NOx concentrations as output.

[0034] It should be noted that the gray-box model has the advantage over the pure mechanistic model in that it does not require a precise mathematical description of all physicochemical processes, thus reducing modeling complexity and computational burden. Compared to the pure black-box model, it has a stronger extrapolation capability, providing reasonable predictions consistent with physical laws even under untrained operating conditions. To enhance the robustness and adaptability of the model, a recursive least squares algorithm with a forgetting factor can be used to update the model parameters online, enabling the predictor to track emission characteristic drift caused by factors such as engine aging and changes in fuel quality.

[0035] For example, taking a certain type of marine low-speed two-stroke diesel engine as an example, the input parameters of the engine source nitrogen oxide predictor include engine speed, fuel supply per cycle, scavenging pressure, scavenging temperature, and exhaust gas recirculation valve opening. The mechanistic model framework can adopt a simplified form of the Zeldowicz mechanism, expressing the nitrogen oxide generation rate as a function of in-cylinder combustion temperature, oxygen concentration, and high-temperature residence time. Using test data of the engine under steady-state conditions at 25%, 50%, 75%, and 100% load, as well as under transient conditions with a step change in load, the gain coefficient, time constant, and pure delay time in the model are identified using a recursive least squares algorithm. The identification results show that the absolute value of the prediction error of the predictor under steady-state conditions is less than 5%, and it can accurately track the actual emission change trend within 2 seconds after a step change in load, providing feedforward information of approximately 1.5 to 3 seconds for subsequent urea injection control, effectively compensating for the response lag caused by exhaust transmission and urea hydrolysis processes.

[0036] S2. Construct and execute a virtual ammonia coverage observer based on extended Kalman filtering to estimate the ammonia coverage on the catalyst surface in the selective catalytic reduction reactor in real time based on the inlet and outlet sensor signals of the reactor.

[0037] The Extended Kalman Filter (EKF) is an extension of the Standard Kalman Filter (SKF). The SKF is suitable for linear Gaussian systems, recursively estimating the system state through two steps: prediction and update. The core idea of ​​the EKF is to perform a Taylor expansion of the nonlinear system at the current estimate, retaining the first-order terms, thus approximating the nonlinear system as linear, and then applying the recursive framework of the SKF. The EKF effectively fuses noisy sensor measurements with system dynamic model information, providing an optimal state estimate while suppressing measurement noise.

[0038] The reason for constructing a virtual ammonia coverage observer is that the ammonia coverage on the catalyst surface cannot be directly measured by physical sensors. Ammonia coverage refers to the proportion of ammonia molecules occupying the active sites of the catalyst, and it is a key variable describing the internal chemical state of a selective catalytic reduction reactor. This variable determines the remaining reaction capacity of the catalyst and the risk of ammonia escape: if the coverage is too low, the denitrification efficiency will be insufficient; if the coverage is too high, excess ammonia will desorb from the catalyst surface and be released into the atmosphere, causing secondary pollution. Since the inside of the catalyst is a closed, high-temperature reaction environment, it is impossible to install measuring equipment. Therefore, ammonia coverage must be indirectly estimated by using measurable inlet and outlet signals such as upstream and downstream nitrogen oxide concentrations, upstream and downstream ammonia concentrations, and exhaust temperature, with the help of state observation algorithms. This alternative measurement method based on software algorithms is called a virtual sensor or soft measurement technology.

[0039] In some implementations, the construction of a virtual ammonia coverage observer can begin with establishing a control-oriented state-space model of a selective catalytic reduction (SCR) reactor. This model is typically based on the principles of material conservation and reaction kinetics, dividing the reactor along the axial direction into multiple continuous stirred tank units, each assuming a uniform distribution of gas component concentrations and catalyst surface coverage. The model's chemical mechanisms include ammonia adsorption and desorption reactions, standard selective catalytic reduction (SCR), rapid selective catalytic reduction (CCR), and ammonia oxidation. These reaction kinetic equations allow for the establishment of a system of nonlinear differential equations with ammonia coverage, nitrogen oxide concentration, and ammonia concentration as state variables.

[0040] After obtaining the state-space model, the extended Kalman filter algorithm is applied to this nonlinear system. Within each sampling period, the algorithm first predicts the state based on the state estimate from the previous time step and the system model, obtaining a prior estimate of the current state. Then, it uses current sensor measurements, such as the reactor outlet nitrogen oxide concentration and outlet ammonia concentration, to calculate the residual between the actual output and the predicted output. Finally, it weights and corrects the prior estimate based on the Kalman gain to obtain a posterior estimate of the current state, where the posterior estimate of ammonia coverage is the output of the virtual observer.

[0041] It should be noted that when applying extended Kalman filtering to selective catalytic reduction systems, special attention must be paid to the convergence and stability of the filter. Since the time constant of ammonia coverage variation is typically on the order of tens of seconds, while the sensor sampling timescale is on the order of seconds to sub-seconds, this timescale separation characteristic is beneficial to the stable operation of the observer. Furthermore, to cope with the wide range of drastic changes in ship engine operating conditions, an adaptive extended Kalman filtering scheme can be adopted. This involves adjusting the process noise covariance matrix and the measurement noise covariance matrix in real time according to changes in operating conditions to ensure the estimation accuracy of the observer across the entire operating range.

[0042] For example, a three-state extended Kalman filter observer is constructed using a selective catalytic reduction system (SCR) for a certain type of ship as the object. Its state vector includes the ammonia coverage on the catalyst surface, the gaseous ammonia concentration in the reactor, and the nitrogen oxide (NOx) concentration in the reactor. The state transition equation is constructed based on material conservation: the rate of change of ammonia coverage is composed of the ammonia adsorption term minus the desorption term and the consumption term participating in the reaction; the rate of change of gaseous ammonia concentration is composed of the ammonia introduced at the inlet minus the ammonia consumed by adsorption plus the ammonia released by desorption and the ammonia carried away at the outlet; the rate of change of NOx concentration is composed of the NOx introduced at the inlet minus the NOx consumed in the reaction and the NOx carried away at the outlet. The measurement vector includes the NOx concentration and ammonia concentration at the reactor outlet. This observer uses the inlet NOx sensor signal and the inlet ammonia sensor signal as model inputs and the outlet sensor signal as the calibration basis.

[0043] S3. Construct and execute a model predictive controller based on ammonia coverage constraints. Use feedforward information as reference input, ammonia coverage on catalyst surface as state feedback, and suppress ammonia escape as constraint objective. Calculate urea injection command through online rolling optimization.

[0044] Ammonia coverage constraint refers to the control requirement of limiting the ammonia coverage on the catalyst surface within a certain optimal range or threshold. Ammonia coverage is directly related to denitrification efficiency and ammonia slip: when ammonia coverage is high, sufficient ammonia is adsorbed on the catalyst surface, which is conducive to the occurrence of nitrogen oxide reduction reactions and thus achieves high denitrification efficiency. However, at the same time, the physically adsorbed ammonia on the catalyst surface is more likely to desorb into the gas phase when the exhaust temperature fluctuates, causing downstream ammonia slip to exceed the limit. Conversely, when ammonia coverage is low, the risk of ammonia slip is reduced, but the denitrification efficiency is insufficient. Therefore, there exists a trade-off range, usually called the optimal ammonia coverage operating range, within which both denitrification efficiency and ammonia slip can be controlled below regulatory limits. The ammonia coverage constraint aims to always limit the system state within this trade-off range.

[0045] Model predictive controllers (MDCs) are a process control strategy whose core idea can be summarized in three steps: In each control cycle, based on the current system state and the predictive model, the trajectory of the system output within a finite time domain is predicted; based on this, an optimization problem within the finite time domain is solved to find the control sequence that optimizes the preset performance indicators; finally, only the first element of the control sequence is applied to the controlled object, and the above process is repeated in the next cycle, forming a rolling time domain or rolling optimization mechanism. The core advantage of MDCs lies in their ability to explicitly handle state constraints and input constraints, directly embedding coverage constraints into the optimization problem, and anticipating future disturbance trends within the prediction time domain, thus achieving forward-looking control.

[0046] In this method, the role of the model predictive controller is reflected in three aspects: First, the feedforward information, namely the predicted source nitrogen oxide concentration, is used as a reference input, enabling the controller to know in advance of the impending disturbance and pre-adjust the urea injection amount; Second, the estimated ammonia coverage is used as state feedback to form a closed-loop correction mechanism to eliminate the deviation caused by model error and unmeasurable disturbance; Third, with the constraint objective of suppressing ammonia escape, the ammonia coverage is ensured to never exceed the safety threshold during the optimization process, thus ensuring emission compliance from the control algorithm level.

[0047] In some implementations, the construction of a model predictive controller can include three main steps: predictive model establishment, optimization problem construction, and rolling optimization solution. Regarding predictive model establishment, a simplified control-oriented model of a selective catalytic reduction system can be used as the predictive model. This model typically simplifies the reactor into one or two continuous stirred tank units, using ammonia coverage as the state variable, urea injection rate as the input variable, and outlet nitrogen oxide concentration and ammonia slip as output variables. Model parameters such as catalyst activity and adsorption-desorption rate constants can be calibrated using bench test data. Regarding optimization problem construction, at each sampling time, the optimization problem that the controller needs to solve is expressed as: under the premise of satisfying the constraints of the ammonia coverage dynamic equation, the upper and lower bound constraints of ammonia coverage, and the upper and lower bound constraints of urea injection rate, find the optimal urea injection rate sequence in the future control time domain that minimizes the objective function value. The objective function typically includes three terms: the sum of squares of the deviations between the outlet nitrogen oxide concentration and the setpoint in the future predicted time domain, characterizing the degree to which the denitrification efficiency meets the standard; the sum of squares of the rate of change of urea injection volume, characterizing the smoothness of the actuator's operation; and the sum of squares of the deviations between the ammonia coverage and the target reference value, characterizing the degree of deviation from the state. Regarding the rolling optimization solution, a quadratic programming algorithm can be used to solve the above constrained optimization problem online. After the solution is completed in each control cycle, the controller outputs the first value of the optimal injection volume sequence to the actuator, and then re-predicts and optimizes in the next cycle, forming a rolling update mechanism.

[0048] It should be noted that the performance of a model predictive controller largely depends on the accuracy of the predictive model and the parameter tuning of the optimization problem. To improve the robustness of the controller to model mismatch, an output deviation correction stage can be introduced into the controller structure, that is, using the deviation between the current model prediction value and the actual measured value to correct the future predicted output. Furthermore, to adapt to the drift of system dynamic characteristics caused by changes in ship engine operating conditions, a scheme combining online model parameter identification and model predictive controller parameter self-tuning can be adopted. Considering the significant asymmetric characteristics of ammonia coverage safety constraints: excessive ammonia coverage leading to ammonia escape violating emission regulations has more serious consequences than the temporary decrease in denitrification efficiency caused by excessively low ammonia coverage. In the objective function and constraint design, asymmetric weighting factors or soft constraint strategies can be used to enable the controller to actively adjust its control behavior as it approaches the constraint boundary.

[0049] S4. Based on the urea injection command, control the urea injection device to inject urea into the exhaust gas to carry out a catalytic reduction reaction of nitrogen oxides in the ship's exhaust gas.

[0050] This step translates the urea injection command calculated in the previous step into actual execution. A urea injection device typically includes components such as a urea storage tank, a urea supply pump, a metering unit, and atomizing nozzles. The urea injection command is the numerical value output by the model prediction controller in the previous step, representing the mass or volume of urea solution to be injected into the exhaust pipe per unit time. After urea is injected into the high-temperature exhaust gas, it first undergoes pyrolysis and hydrolysis: when the exhaust temperature exceeds 160 degrees Celsius, urea begins to decompose into ammonia and isocyanate. The isocyanate further reacts with water vapor to produce ammonia and carbon dioxide. The generated ammonia diffuses to the catalyst surface under the influence of airflow, is adsorbed by active sites to form ammonia coverage, and then undergoes a selective catalytic reduction reaction with nitrogen oxides diffused to the catalyst surface to produce harmless nitrogen and water vapor.

[0051] In some implementations, the urea injection device can be controlled using an electronically controlled injection scheme based on pulse width modulation (PWM) signals. Specifically, a high-precision solenoid valve or piezoelectric injection valve is integrated into the urea metering unit. The controller calculates the required valve opening duration based on the urea injection command and outputs a PWM drive signal with the corresponding duty cycle. The amount of urea injected per unit PWM cycle is pre-calibrated using the flow characteristic curve of the injection valve. During actual injection, feedforward compensation can be performed on the injection quantity based on upstream exhaust flow and exhaust temperature signals. To ensure injection accuracy, a closed-loop control strategy can be adopted: a pressure sensor is installed at the urea supply pump outlet, and the pump speed or return valve opening is adjusted using a proportional-integral-derivative (PID) controller to stabilize the injection pressure at a set value, such as 5 bar to 9 bar; simultaneously, the actual injection quantity is accurately calculated using the pressure difference signal before and after the metering valve and the valve opening time. In addition, to prevent urea crystallization and nozzle blockage, a purging process must be performed before the engine is stopped, using compressed air to purge the residual urea solution from the pipeline and nozzles, and a cooling water jacket is installed at the nozzle to control the nozzle temperature within a safe range.

[0052] It should be noted that the execution quality of urea injection directly affects the performance of the entire denitrification system. Insufficient urea injection volume leads to insufficient ammonia and substandard denitrification efficiency; excessive injection volume results in excess ammonia, causing ammonia escape and pollution, and the excess urea may not decompose completely, forming solid deposits such as biuret and melamine, which can clog catalyst channels or nozzles. Therefore, the execution process should possess fault diagnosis and tolerance capabilities. When low urea level, abnormal supply pump pressure, nozzle blockage, or drive circuit failure is detected, the control system should be able to identify the fault type and take appropriate safety measures, including but not limited to switching to redundant injection units, limiting engine power output, and activating alarm indicator lights. Simultaneously, fault information should be recorded and reported to the engine room personnel or shore-based monitoring center via the shipboard communication interface.

[0053] Based on the above technical solutions, the intelligent control method for ship exhaust gas denitrification provided in this application overcomes the response lag problem of selective catalytic reduction systems by generating feedforward information through an engine-source nitrogen oxide predictor; it achieves real-time estimation of the internal chemical state of the catalyst through a virtual ammonia coverage observer, solving the engineering problem of unmeasurable key variables; and through an ammonia coverage-constrained model predictive controller, the ammonia coverage is strictly controlled within a trade-off range under the synergistic effect of feedforward information and state feedback, achieving a balance optimization between denitrification efficiency and ammonia slip. The organic combination of these three elements constitutes a complete intelligent control closed loop from feedforward prediction and state perception to constraint optimization.

[0054] Furthermore, in the embodiments of this application, the above control method can be further extended to the following schemes: integrating ship navigation plan information and GPS data into the gray box model to predict the load change trend of future voyages in advance and achieve advanced predictive control; expanding the virtual ammonia coverage observer into a multi-state observer to simultaneously estimate health status parameters such as catalyst aging degree and sulfur poisoning degree, providing a basis for maintenance decisions; expanding the single-level model predictive controller into a cascaded or distributed model predictive control architecture to control the ammonia coverage of upstream and downstream units respectively, further improving control accuracy; introducing an injection quantity correction module based on catalyst temperature to perform low-temperature compensation or high-temperature suppression of urea injection commands according to the inlet temperature of the selective catalytic reduction reactor, in order to address the problem of insufficient urea hydrolysis caused by low exhaust temperature when the ship is running at low load.

[0055] In one possible implementation of the embodiments of this application, combined with Figure 2 The above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below: S101. Constructing a mechanistic model based on the extended Zeldovich mechanism involves calculating the in-cylinder characteristic temperature using combustion characteristic parameters from the engine's real-time operating parameters, and calculating the nitrogen oxide generation rate based on the Arrhenius equation to obtain the mechanistic model output value.

[0056] The extended Zeldovich mechanism is a classical chemical kinetic theory describing the formation of thermodynamic nitrogen oxides (NOx). This theory posits that NOx formation primarily depends on three elementary reactions: ; ; ; In the above reactions, R1 has the highest activation energy and is the rate-determining step. The total rate of nitrogen oxide formation is determined by the rates of all three reactions, and can be expressed as: The rate constants for each reaction all obey the Arrhenius equation: In the above formula, T is the absolute temperature of the reaction, in Kelvin; R is the ideal gas constant, with a value of 8.314 joules per mole per Kelvin. Pre-exponential factor, Temperature index These parameters, which are activation energies, can be obtained from chemical kinetic databases or engine bench calibrations.

[0057] In marine diesel engines, the combustion temperature inside the cylinder is uneven, and characteristic temperatures are typically used to address this. This temperature range is used to approximate the main temperature range for NOx formation. This characteristic temperature can be estimated using combustion characteristic parameters from real-time engine operating parameters, such as calculating the air-fuel ratio based on fuel injection quantity and intake air quantity, estimating the temperature rise caused by combustion heat release according to the first law of thermodynamics, and finally superimposing it with the intake air temperature to obtain the final result. ; in, Intake air temperature, For combustion efficiency, Because it is a low-calorific-value fuel, For the mass of fuel injected per cycle, The specific heat capacity of exhaust gas at constant pressure. This refers to the intake air mass per cycle.

[0058] To derive the concentration of source nitrogen oxides in the exhaust gas from the nitrogen oxide formation rate, integration over time is required. Under steady-state conditions, a constant formation rate can be assumed, meaning the mass of NOx generated per cycle is inversely proportional to engine speed. The final output value of the mechanistic model is obtained as follows: ;in, The molar mass of nitric oxide is... For exhaust gas density, Where N is the exhaust volume for a single cycle and N is the engine speed. This refers to the duration of combustion.

[0059] S102. Construct a black-box correction model based on radial basis function network. Take the real-time operating parameters of the engine as input and output the deviation correction value between the output value of the mechanism model and the actual source nitrogen oxide concentration. This is used to compensate for the prediction deviation caused by the non-ideal combustion of the mechanism model.

[0060] The Radial Basis Function Network (RBF) is a feedforward neural network with a single hidden layer, where the hidden neurons use radial basis functions as activation functions. This network can approximate continuous nonlinear functions with arbitrary precision, has fast training speed, and does not suffer from local minima. The black-box correction model part learns complex nonlinear relationships that the mechanistic model fails to capture, such as uneven in-cylinder temperature distribution, local oxygen deficiency in the combustion chamber, the effects of exhaust gas recirculation, and variations in injection timing, thereby outputting a correction term. , so that: Approaching the measured value .

[0061] The mathematical expression for a radial basis function network is: ;in, The input vector includes engine speed, load, fuel injection quantity, boost pressure, exhaust temperature, exhaust flow rate, and excess air coefficient; M is the number of hidden layer neurons. Let j be the center vector of the j-th neuron; For width parameters; These are the output layer weights. Parameters to be identified. Including all , and .

[0062] S103. Perform offline identification of the black-box correction model part with the goal of minimizing the root mean square error between the predicted value and the measured value, and introduce a regularization term to prevent overfitting and ensure the model's extrapolation ability outside the training data domain.

[0063] The regularization term includes a weight decay term and a gradient constraint term. The gradient constraint term is used to force the gradient of the black box correction model part with respect to the engine operating parameters to remain orthogonal to the gradient of the mechanism model part with respect to the engine operating parameters, so that the black box correction model part only compensates for the nonlinear residuals not captured by the mechanism model part, without changing the main physical law of nitrogen oxide generation changing with operating conditions reflected by the mechanism model part.

[0064] The core of offline identification is solving the following constrained nonlinear least squares problem: ; The first term is the sum of squared prediction errors, ensuring model fitting accuracy; the second term is the weight decay regularization term. The first term is the weight decay coefficient, used to prevent overfitting; the second term is the gradient constraint regularization term. These are the gradient constraint coefficients. and These are the gradient vectors of the black-box correction model and the mechanistic model with respect to the engine operating parameters, respectively.

[0065] The physical significance of gradient orthogonality constraints lies in the fact that if the mechanistic model has correctly described the main trend of the source NOx concentration increasing with engine speed, that is... Therefore, the black-box correction model should no longer change this trend, that is... It should approach zero. By minimizing the squared L2 norm of the gradient difference, the black-box correction model is forced to compensate only for the higher-order nonlinear residuals not captured by the mechanistic model, without changing the main physical laws, thus ensuring the extrapolation capability of the gray-box model outside the training data domain.

[0066] In terms of algorithm design, the above optimization problem can be solved using gradient descent, the Levenberg-Marquardt algorithm, or the Adam optimizer. Since the objective function includes a gradient term, calculating the gradient requires the second derivative of the black-box model, which can be achieved using automatic differentiation techniques or numerical difference approximation. To improve convergence speed, it is recommended to first use conventional training without gradient constraints (i.e.,...). =0) to obtain the initial parameters, and then gradually increase them. Make minor adjustments.

[0067] Finally, by adding the output value of the mechanistic model to the output value of the black-box correction model obtained through offline identification, the predicted value of the source nitrogen oxide concentration is obtained, and this predicted value is output as feedforward information to the model prediction controller.

[0068] Based on the above technical solution, this application constructs an engine-source nitrogen oxide predictor using a gray-box modeling approach. This predictor integrates the advantages of physical mechanisms and data-driven approaches, and utilizes gradient orthogonal constraints to ensure the physical consistency and extrapolation capability of the model. In Simulink, this predictor can be implemented as a parallel addition of a mechanism model module and an RBF correction module, providing accurate feedforward information for the subsequent model predictive controller.

[0069] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Establish a discrete state-space model of the selective catalytic reduction reactor, with the ammonia coverage on the catalyst surface and the gaseous ammonia concentration in the selective catalytic reduction reactor as state variables, the ammonia concentration corresponding to the urea injection rate as the control input, and the exhaust temperature and exhaust flow rate as time-varying parameters, and construct the state transition equation.

[0070] The physicochemical processes occurring within the selective catalytic reduction (SCR) reactor mainly include: adsorption of ammonia (NH3) from the gas phase onto the catalyst surface, desorption of adsorbed ammonia, and the catalytic reduction reaction of adsorbed ammonia with gaseous nitrogen oxides (NOx). To describe these processes, a classic one-dimensional two-state model can be used, with the state variables being the ammonia coverage θ on the catalyst surface and the gaseous ammonia concentration. Ammonia coverage θ is defined as the ratio of occupied active sites to the total number of active sites, ranging from 0 to 1. It is a core state determining denitrification efficiency and ammonia escape risk. (Gaseous ammonia concentration) This indicates the concentration of unadsorbed ammonia gas in the SCR reactor, expressed in moles per cubic meter.

[0071] In some implementations, the continuous-time form of the state transition equation is determined by three reaction rates: adsorption rate Desorption rate SCR reaction rate .in , , All are rate constants, obey the Arrhenius equation, and are functions of the exhaust temperature T; The concentration of gaseous nitrogen oxides entering the SCR reactor is given by feedforward information from the engine-source nitrogen oxide predictor. This leads to the differential equations for the state variables: ; ; in, This refers to the exhaust mass flow rate. For exhaust density, For catalyst volume, This refers to the ammonia concentration at the SCR inlet (i.e., the ammonia concentration produced after urea injection pyrolysis). The apparent density of the catalyst coating, For saturated ammonia coverage, The value represents the porosity of the catalyst.

[0072] To facilitate the implementation of the digital controller, the above continuous model needs to be discretized. A first-order Euler discretization method is used, with a sampling period of... The discrete state transition equations are obtained as follows: ; ; It should be noted that the kinetic parameters of the SCR reactor are as follows: , , The parameters typically change with catalyst aging, therefore, periodic online parameter identification or adaptive updates are necessary in practical applications. Additionally, the discretization sampling period... The choice of value is crucial: too small a value will lead to excessive computational burden, while too large a value may result in the loss of system dynamic characteristics. Generally, a value of is recommended. Between 0.1 seconds and 1 second.

[0073] S202. Construct an observation equation based on the cross-sensitivity characteristics of the nitrogen oxide sensor, and adopt a dual-time-scale dynamic separation model to express the measurement signal of the nitrogen oxide sensor as the superposition of the actual nitrogen oxide concentration response and the actual ammonia concentration response at the outlet.

[0074] The dynamic characteristics of the actual nitrogen oxide concentration response are characterized by a first time constant, and the dynamic characteristics of the actual ammonia concentration response are characterized by a second time constant, with the second time constant being smaller than the first time constant.

[0075] The nitrogen oxide sensor is a key downstream measurement device in the SCR system, but its output signal S(t) is not only sensitive to NOx but also exhibits cross-sensitivity to NH3. This means that when ammonia is present in the exhaust gas, the sensor reading will be artificially high. Furthermore, the sensor responds differently to NOx and NH3, typically responding faster to NH3. To separate the true NOx and NH3 concentrations from the sensor signal for state observation, an observation equation reflecting this dynamic cross-sensitivity characteristic must be constructed.

[0076] In some implementations, the observation equations in the continuous-time domain are expressed in convolutional form: ; Where S(t) is the measurement signal from the nitrogen oxide sensor, This represents the actual nitrogen oxide concentration at the SCR reactor outlet. This represents the actual ammonia concentration at the export price. The cross sensitivity coefficient (typically between 0.8 and 1.2) is used. The sensor's response time constant to nitrogen oxides is... Let be the sensor's response time constant to ammonia, and v(t) represents the measurement noise, and the symbol denotes the convolution operation. This convolution form essentially describes the step response characteristics of a first-order linear system.

[0077] To implement this in a discrete-time control system, the above continuous observation equations are discretized, resulting in: ; in, , , The sampling period is This is the recursive term of the ammonia response at the previous moment (i.e., the part of the sensor signal contributed by ammonia at the previous moment). The sensor measurement signal at the current moment. The sensor measurement signal from the previous moment. and These are the actual NOx and NH3 concentrations at the current export time. This is for discrete measurement noise.

[0078] It should be noted that in the observation equation, the actual NOx concentration at the outlet... With state variables and The relevant, specific relationship is as follows: ,in For space and time; while the actual NH3 concentration at the outlet. This is approximately equal to the gaseous ammonia concentration in the SCR reactor. (Assuming homogeneous mixing within the reactor). Therefore, the observation equation establishes the measurable sensor signal. State variables that cannot be directly measured , The connection between them.

[0079] S203. Based on the state transition equation and observation equation, perform extended Kalman filter recursion to output the estimated value of ammonia coverage on the catalyst surface in real time, and output the estimated value as a state feedback signal to the model predictive controller.

[0080] The Extended Kalman Filter (EKF) is a recursive filtering algorithm used for state estimation of nonlinear systems. Since both the state transition equation and the observation equation of an SCR system are nonlinear (the state transition equation contains...),... Since the nonlinear function has terms and an exponential function, the EKF must be used instead of the standard Kalman filter. The core idea of ​​the EKF is to perform a first-order Taylor linearization of the nonlinear function near the current estimate, and then apply the prediction update framework of the standard Kalman filter.

[0081] In some implementations, the recursive steps of the Extended Kalman Filter (EPF) are divided into two stages: prediction and update. First, the state vector is defined. Control input Known disturbance The time-varying parameters include exhaust temperature T_k and exhaust flow rate. The nonlinear state transition function is denoted as... The nonlinear observation function is denoted as ,in can be and Calculated.

[0082] Prediction phase: Estimating the state from the previous time step And control input calculation one-step prediction state Simultaneously calculate the state transition Jacobian matrix. Then update the error covariance matrix. ,in Let be the process noise covariance matrix.

[0083] Update phase: First, calculate the observed Jacobian matrix. Then calculate the Kalman gain. Where R is the measurement noise variance. Obtain the sensor measurement value at the current moment. (Right now ), calculate new information Last updated state estimate And update the error covariance matrix. .

[0084] It should be noted that the performance of the extended Kalman filter depends on the process noise covariance. and measurement noise covariance A reasonable selection. This reflects the uncertainty of the model. This reflects the noise level of the sensor. Tuning can typically be done through offline experiments or trial and error. Additionally, to ensure filter stability, the state transition Jacobian matrix needs to be guaranteed. and the observed Jacobian matrix It is computable and nonsingular at every step.

[0085] For example, setting the sampling period =0.2 seconds, process noise covariance Here, `diag()` represents constructing a diagonal matrix, which is a diagonal function in matrix operations; the measurement noise covariance R = 0.1. The initial state estimate is set to... =0.2, Initial covariance matrix In each control cycle, the module reads the ammonia concentration corresponding to the current urea injection. Exhaust temperature T, exhaust flow rate Feedforward information and downstream NOx sensor measurements An extended Kalman filter recursion is performed to obtain an estimate of the ammonia coverage on the catalyst surface. This estimate is sent to the model prediction controller in real time as status feedback.

[0086] Based on the above technical solution, this application achieves the function of estimating the ammonia coverage on the catalyst surface in real time without additional hardware by constructing a discrete state-space model, introducing observation equations with dynamic separation of dual time scales, and performing extended Kalman filter recursion. This virtual observer fully utilizes the cross-sensitivity characteristics of existing NOx sensors, overcomes the problem of unobservable state in traditional methods, and provides key state feedback information for the model predictive controller.

[0087] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Establish a simplified ammonia storage kinetic model for predictive control. Use a first-order equation of state to describe the dynamic change of ammonia coverage on the catalyst surface. Use the ammonia concentration corresponding to the urea injection command as the control input, the source nitrogen oxide concentration corresponding to the feedforward information as the known disturbance input, and the estimated value of ammonia coverage on the catalyst surface output by the virtual ammonia coverage observer as the current value of the state variable.

[0088] The Model Predictive Controller (MPC) needs to predict future system behavior based on a simplified predictive model. To reduce online computational complexity while retaining key dynamics affecting control performance, a first-order ammonia storage kinetic model is used as the predictive model. This model ignores the dynamic details of gaseous ammonia concentration and focuses only on the change in ammonia coverage θ on the catalyst surface, as θ directly determines denitrification efficiency and ammonia slip risk. The simplified model assumes that the gaseous ammonia concentration is proportional to the injected ammonia concentration and that the adsorption, desorption, and reaction processes reach a quasi-steady state.

[0089] In some implementations, the simplified continuous-time form of the ammonia storage kinetic model is as follows: ; in, The ammonia coverage on the catalyst surface is dimensionless. The concentration of ammonia gas produced after urea injection pyrolysis is the control input u; The concentration of source nitrogen oxides entering the SCR reactor is given by feedforward information provided by the engine source nitrogen oxide predictor as a known disturbance input; , , Let be the adsorption, desorption, and reaction rate constants, respectively, all of which are functions of the exhaust temperature T and obey the Arrhenius equation: , , ; To achieve prediction in a digital controller, the continuous model described above needs to be discretized. A first-order Euler discretization method is used, with a sampling period of... (In accordance with the observer and controller), a discrete prediction model is obtained: ; in, To control the input, Given a known disturbance, The exhaust temperature is at the current moment. Although this first-order model is simple, it can capture the integral effect of ammonia coverage on the amount of ammonia injected, as well as the effect of temperature on the reaction rate, making it sufficient for rolling optimization of MPC.

[0090] It should be noted that the dynamic parameters in the prediction model , , The parameters should be consistent with those used by the virtual ammonia coverage observer to ensure consistency between state feedback and model predictions. If catalyst aging causes parameter drift, it can be corrected through online identification.

[0091] S302. Construct a constraint processing module. In each control cycle, dynamically calculate the upper limit constraint value of ammonia coverage on the catalyst surface based on the currently measured exhaust temperature and space velocity, and set control input amplitude constraints.

[0092] The constraint processing module is one of the key features that distinguishes the model predictive controller from traditional PID control. The main cause of ammonia escape is excessive ammonia coverage on the catalyst surface, leading to excessive ammonia desorption and discharge with the exhaust gas. However, the maximum allowable ammonia coverage is not fixed: as exhaust temperature increases, the ammonia desorption rate accelerates, reducing the amount of ammonia the catalyst can safely store, thus the upper bound should be decreased; conversely, at low temperatures, ammonia is less prone to desorption, so the upper bound can be appropriately increased. Simultaneously, high space velocity means a short residence time of the exhaust gas within the catalyst, resulting in insufficient ammonia adsorption, which also necessitates a lower upper bound. Therefore, a dynamic upper bound constraint function must be established.

[0093] In some implementations, the formula for calculating the upper bound constraint value of dynamic ammonia coverage is: ;in, This represents the maximum permissible ammonia coverage under current operating conditions. is the saturated ammonia coverage (usually taken as 0.95); T is the exhaust temperature; Airspeed, This refers to the exhaust volume flow rate. This represents the catalyst volume; The saturated gas phase ammonia concentration (determined by the sensor range or catalyst characteristics). and These are the rate constants for adsorption and desorption reactions, respectively. The physical meaning of this formula is: under steady state, the ammonia coverage and the gaseous ammonia concentration satisfy the Langmuir isotherm adsorption relationship; when the gaseous ammonia concentration reaches saturation... The corresponding ammonia coverage at that time is the upper bound. By introducing the reciprocal of the space velocity, the effect of residence time on the adsorption equilibrium is reflected.

[0094] In each control cycle, based on the currently measured exhaust temperature and airspeed Real-time calculation of dynamic upper bound: The upper bound increases with increasing exhaust temperature and decreases with increasing airspeed. Simultaneously, control input amplitude constraints are set: ,in It is usually 0. The maximum flow rate is determined by the urea injection system. Additionally, constraints on the rate of change of control input can be added as needed.

[0095] It should be noted that the calculation of the dynamic upper bound constraint value depends on the adsorption and desorption rate constants, which can be obtained through catalyst sample testing or vehicle calibration. In practical applications, due to catalyst aging, and These parameters will change, so they need to be updated periodically. Additionally, The calculation should be completed within each sampling period. Its computational load is small and will not burden the MPC solution.

[0096] S303. Construct an optimization solution module with the goal of minimizing the sum of squares of the deviation between the predicted outlet nitrogen oxide concentration and the target value in the time domain. With the constraints of the control input amplitude and the dynamic upper bound constraint of the ammonia coverage on the catalyst surface as constraints, construct and solve the rolling time domain optimization problem. The first element of the optimal control sequence obtained by the solution is used as the urea injection command output at the current moment.

[0097] The core of model predictive control is to solve a finite-time open-loop optimal control problem at each sampling time, apply the calculated first control action to the system, and then repeat the process at the next time step; this is called rolling time-domain control. Optimization objectives typically include output tracking deviation and penalties for changes in control inputs to ensure denitrification efficiency and control stability. Constraints protect against ammonia slip risk and ensure hardware safety.

[0098] In some implementations, the prediction time domain length is The control time domain length is (generally The outlet nitrogen oxide concentration at step i in the prediction time domain is defined as... This value can be calculated using a prediction model and current state estimation. Specifically, the relationship between the outlet concentration and ammonia coverage is as follows: ;in The source NOx concentration provided for the feedforward information. For spacetime.

[0099] The mathematical formulation of the optimization problem is as follows: ; ; ; ; in, For control sequences; The target NOx concentration at the outlet (usually the limit to meet emission regulations); To control the incremental weighting coefficient, which is used to suppress drastic fluctuations in the injection volume; ; The upper bound of the ammonia coverage predicted in step i is calculated based on the predicted exhaust temperature and space velocity.

[0100] Since the prediction model is nonlinear, the above optimization problem belongs to the nonlinear model predictive control (NMPC) problem. Solution methods can include sequential quadratic programming (SQP), interior-point methods, or direct multi-shot methods. To reduce the online computational burden, the nonlinear model can also be linearized at each sampling time point, transforming it into a quadratic programming (QP) problem, and then solved using an efficient QP solver (such as qpOASES or OSQP). The linearized prediction model is as follows: ; where the coefficient , , From the current state It is determined by the operating parameters.

[0101] The optimal control sequence is obtained by solving the problem. Then, only the first element This serves as the output of the urea injection command at the current moment, i.e., the ammonia concentration command. In the next sampling cycle, the optimization problem is resolved based on the new state estimate and operating condition information.

[0102] It should be noted that real-time computation is a key challenge in rolling time-domain optimization. The sampling period of a ship's engine controller is typically 0.1 to 0.5 seconds, therefore the optimization solution must be completed within this sampling period. Using linearized MPC combined with an efficient QP solver can meet the real-time requirements. For nonlinear MPC, a shorter prediction time domain (such as...) can be used. =5、 =2) or use explicit MPC to pre-compute the piecewise affine form of the optimal solution.

[0103] For example, setting the sampling period =0.2 seconds, prediction time domain =10, control time domain =3, target outlet NOx concentration =200 mg / m³, controlling incremental weighting =0.01. At the current moment, the virtual observer outputs an estimated ammonia coverage value. =0.65, feedforward information source NOx concentration =600 mg / m³, exhaust temperature =320 degrees Celsius, airspeed =30,000 per hour. The constraint processing module calculates the dynamic upper bound. =0.82. The MPC optimization solver solves the above nonlinear optimization problem online to obtain the optimal ammonia concentration corresponding to urea injection. =150 mg / m³, which is converted into a urea injection pulse width command and sent to the injection device. The process is repeated at the next moment.

[0104] Based on the above technical solutions, this application achieves precise control of urea injection by establishing a simplified ammonia storage kinetic prediction model, introducing a dynamic ammonia coverage upper bound constraint, and constructing a rolling time-domain optimization solver. This model predictive controller can fully utilize feedforward information to predict disturbances, use state feedback to avoid blind ammonia injection, and actively suppress ammonia escape while ensuring denitrification efficiency, significantly improving the control performance of the ship's exhaust gas denitrification system under complex operating conditions.

[0105] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S4 specifically includes the following S401 to S403: S401: Receive the urea injection command output by the model prediction controller and convert the urea injection command into a drive control signal for the urea injection device.

[0106] The urea injection command output by the model predictive controller (MPC) is typically expressed as an ammonia concentration. The ammonia concentration is expressed in milligrams per cubic meter or moles per cubic meter. This instruction indicates the total amount of ammonia that needs to be injected into the exhaust gas. However, actual ship exhaust aftertreatment systems use an aqueous urea solution with a mass concentration of 32.5% or 40%. Therefore, the ammonia concentration instruction needs to be converted into the mass flow rate or injection pulse width of the aqueous urea solution to drive the metering valve in the urea injection device.

[0107] In some implementations, the conversion formula is based on mass conservation and pyrolysis efficiency. Ammonia is derived from the thermal decomposition reaction of urea. Theoretically, 1 mole of urea can produce 2 moles of ammonia. However, considering incomplete decomposition and side reactions during actual pyrolysis, pyrolysis efficiency is introduced. (Typically between 0.85 and 0.98). Therefore, the required mass flow rate of the urea solution is... With the target ammonia concentration The relationship is: ; in, This refers to the exhaust volumetric flow rate, expressed in cubic meters per hour. =60.06 grams per mole is the molar mass of urea; =17.03 grams per mole is the molar mass of ammonia; This represents the mass concentration of the urea aqueous solution, typically 0.325. After obtaining the mass flow rate, it is further converted into the duty cycle or injection pulse width of the metering valve based on the characteristics of the injection device (such as nozzle flow coefficient, supply pressure, etc.), generating the corresponding drive signal.

[0108] It should be noted that the measurement of exhaust flow rate in the above conversion has a certain lag and error. Therefore, in actual control, it can be estimated in real time by combining exhaust temperature, engine speed, and load. Furthermore, pyrolysis efficiency is closely related to exhaust temperature: when the exhaust temperature is below 250 degrees Celsius, the pyrolysis efficiency decreases significantly, and urea crystals may even form. In this case, it is necessary to adjust the conversion formula... Perform temperature correction.

[0109] S402. The urea injection device is controlled to inject urea aqueous solution into the exhaust pipe according to the drive control signal. The exhaust heat is used to pyrolyze the urea to generate ammonia. The ammonia is mixed with the exhaust gas and then enters the selective catalytic reduction reactor.

[0110] The urea injection system typically consists of a urea tank, a supply pump, a metering valve, and nozzles. Upon receiving a drive control signal, the metering valve opens to a specified degree or pulse width, and the high-pressure urea solution is atomized through the nozzles and injected into the exhaust pipe. The atomized urea droplets rapidly absorb heat, evaporate, and undergo thermal decomposition in the high-temperature exhaust gas, generating gaseous ammonia. Simultaneously, the exhaust flow promotes the mixing of ammonia and tail gas, forming a relatively homogeneous ammonia-exhaust gas mixture, which then enters the downstream selective catalytic reduction (SCR) reactor.

[0111] In some implementations, to improve pyrolysis efficiency and prevent urea crystallization, the nozzle is typically installed in the high-temperature zone of the exhaust pipe, and a mixer or baffles are placed around the nozzle to enhance gas-liquid heat transfer and mixing. The rate of the pyrolysis reaction is affected by the exhaust temperature, droplet diameter, and residence time. To accurately describe the pyrolysis process, a first-order kinetic model can be used: ;in, The mass of urea that has not yet been decomposed. This is the pyrolysis rate constant. This model can be used to estimate the actual effective ammonia concentration entering the SCR reactor, thereby providing feedforward compensation for the injection command.

[0112] It should be noted that under low-temperature conditions where the exhaust temperature is below 200 degrees Celsius, urea pyrolysis is incomplete, easily forming white crystalline deposits on the exhaust pipe wall, mixer, or catalyst inlet, leading to increased back pressure and nozzle blockage. To address this issue, this application also incorporates an active thermal state pre-regulation strategy: when a low-temperature condition is predicted, the heating device is activated in advance to increase the exhaust temperature or catalyst bed temperature, thereby ensuring pyrolysis efficiency and system reliability.

[0113] S403. In the selective catalytic reduction reactor, ammonia is adsorbed on the catalyst surface and undergoes a selective catalytic reduction reaction with nitrogen oxides to generate nitrogen and water. At the same time, according to the active thermal state pre-regulation strategy, the heating device is started in advance when the low temperature condition is predicted to maintain the catalyst bed temperature within the active window.

[0114] The selective catalytic reduction (SCR) reactor is internally coated with a vanadium-based or zeolite-based catalyst, and its core reaction is: The reaction rate is strongly dependent on the ammonia coverage on the catalyst surface. And the reaction temperature T. Under ideal conditions, when the ammonia coverage is controlled between 0.6 and 0.8 and the temperature is in the range of 280 to 450 degrees Celsius, the denitrification efficiency can reach over 90%. In order to address the problem of catalyst activity reduction and ammonium bisulfate ABS condensation caused by excessively low exhaust temperature during low-load operation of marine engines, this application introduces an active thermal state pre-conditioning strategy.

[0115] The active thermal state pre-regulation strategy is based on predicted navigation condition information. It obtains the ship's speed change trajectory, port entry / exit plans, or engine load change trajectory over a future period through the ship's navigation planning system or autopilot system. A pre-established thermal inertia model is used to predict the future trend of catalyst bed temperature changes in the SCR reactor. The thermal inertia model can be described as a first-order inertial element plus a pure delay element: ;in, The thermal time constant, For the delay of exhaust gas transmission, The catalyst bed temperature, The exhaust temperature, For active heating power, This refers to the exhaust mass flow rate. The specific heat capacity of the exhaust gas. When it is predicted that the catalyst bed temperature will fall below a preset low-temperature threshold (e.g., 260 degrees Celsius) within a certain period in the future, the heating device (e.g., burner or electric heater) connected to the SCR reactor is activated in advance at the current moment to preheat the catalyst to the target temperature range (e.g., 280 to 300 degrees Celsius) before entering the low-temperature operating condition. The start time and heating power of this heating action can be determined by solving a minimum energy consumption optimization problem.

[0116] It should be noted that the active thermal preconditioning and the MPC control in S301 to S303 work in tandem: the MPC is responsible for optimizing urea injection under normal operating conditions, while the active preconditioning is responsible for creating suitable reaction conditions for the catalyst before low-temperature operating conditions. The two work together by sharing navigation condition prediction information to avoid a sudden drop in denitrification efficiency and ABS blockage caused by excessively low exhaust temperatures.

[0117] In the Simulink implementation framework, S403 comprises two parallel components: a chemical reaction model for the SCR reactor, used to simulate the denitrification process; and an active thermal state pre-regulation controller. The chemical reaction model can employ the same discrete state-space model as S201, with ammonia and source NOx concentrations as inputs and outlet NOx concentrations and ammonia slip as outputs. The active thermal state pre-regulation controller can be designed as a prediction-based on-off controller or an MPC controller: inputs are predicted navigation trajectory, and outputs are start / stop commands for the heating device and heating power. This module connects to the MPC controller in S3 via a data bus, sharing operating condition information and state estimates.

[0118] For example, the ship plans to enter the emission control zone in 10 minutes, at which time the engine will operate at reduced load, and the exhaust temperature is expected to drop from 350 degrees Celsius to 220 degrees Celsius. The active thermal pre-conditioning controller predicts, based on a thermal inertia model, that without heating, the catalyst bed temperature will drop below 250 degrees Celsius in 12 minutes, below the threshold of 260 degrees Celsius. Therefore, the controller activates the burner at this moment to heat the exhaust gas at 20 kW, maintaining the catalyst bed temperature above 280 degrees Celsius before entering the low-temperature zone, thus ensuring a denitrification efficiency of no less than 85%.

[0119] Based on the above technical solution, this application reliably converts the urea injection command output by the model predictive controller into actual injection actions through the coordinated execution of S401 to S403. Combined with an active thermal state pre-regulation strategy, it effectively solves the problems of catalyst deactivation and urea crystallization under low-temperature conditions. The entire injection-pyrolysis-catalytic reaction process forms a closed loop with the upstream prediction, observation, and control links, ensuring that the ship's exhaust gas denitrification system can maintain a highly efficient, stable, and low-escape operating state across the entire operating range.

[0120] In one possible implementation of this application embodiment, the intelligent control method for ship exhaust gas denitrification further includes active thermal state pre-adjustment, which can be specifically implemented through the following steps S501, S502, and S503: S501. Construct a thermal inertia model for a selective catalytic reduction reactor, using first-order inertia plus a pure delay element to describe the dynamic response characteristics of the catalyst bed temperature as a function of exhaust temperature and active heating power.

[0121] The thermal inertia model is used to quantitatively describe the response characteristics of the catalyst bed temperature in a selective catalytic reduction (SCR) reactor to changes in exhaust temperature and active heating power. Because the SCR reactor has a large heat capacity, its temperature changes exhibit significant lag and inertia relative to changes in exhaust temperature. An accurate thermal inertia model is fundamental for temperature prediction and pre-regulation control.

[0122] In some implementations, the thermal inertia model adopts a first-order inertia plus pure delay transfer function form: ;in, The catalyst bed temperature is expressed in degrees Celsius. Exhaust temperature, in degrees Celsius; This refers to the active heating power, measured in watts. This refers to exhaust mass flow rate, expressed in kilograms per second. is the specific heat capacity of the exhaust gas at constant pressure, in joules per kilogram per Kelvin; K is the steady-state gain (usually close to 1). is the thermal time constant, expressed in seconds, which characterizes the magnitude of the inertia of temperature changes; This is the pure delay time, measured in seconds, which characterizes the time required for heat to transfer from the heating device to the catalyst bed.

[0123] To achieve prediction in the digital controller, the above continuous model is discretized, resulting in a heat dissipation inertial model in the form of difference equations: ;in, , The sampling period; is the delay step number (rounded to an integer); k is the sampling time sequence number.

[0124] It should be noted that the parameters in the thermal inertia model and The parameters are related to the structural dimensions of the SCR reactor, the thermal properties of the materials, and the exhaust flow rate. They can typically be identified through a step response experiment: under steady-state conditions, a step change in exhaust temperature or a step power increase is applied to the heating device, the response curve of the catalyst bed temperature is recorded, and then the parameter values ​​are obtained through curve fitting. For different exhaust flow rate ranges, a lookup table model can be established.

[0125] S502. Based on the predicted navigation conditions and thermal inertia model, the catalyst bed temperature of the selective catalytic reduction reactor is predicted within a preset time period in the future, and the predicted temperature trajectory is obtained.

[0126] The predicted navigation condition information includes the ship's speed change trajectory, port entry and exit plans, or engine load change trajectory within a preset time period. This information can be obtained from the ship's navigation planning system or autopilot system. Based on the mapping relationship between engine load and exhaust temperature (usually obtained through bench calibration), the speed or load trajectory can be converted into an exhaust temperature trajectory. and exhaust flow trajectory Then, using the thermal inertia model established by S501, combined with the measured or estimated values ​​of the catalyst bed temperature at the current moment, the catalyst temperature over a future period is recursively predicted to obtain the predicted temperature trajectory. .

[0127] In some implementations, the prediction algorithm uses a rolling time-domain recursive approach. The prediction duration is set. (e.g., 30 minutes), sampling step size (For example, 1 second). In each control cycle, the catalyst bed temperature at the current moment is obtained. (Obtained through temperature sensor measurement or state estimation). Then, based on the predicted exhaust temperature sequence... and exhaust flow sequence (in Assuming active heating power The catalyst temperature is kept at an undetermined value or temporarily set to 0 for a period of time, and the predicted values ​​for each future time are calculated recursively using the heat dissipation inertia model: , ; where the initial conditions When active heating is not considered, take directly Prediction can be made; if the heating effect needs to be considered to determine the pre-conditioning strategy, an iterative approach can be used to solve the problem.

[0128] It should be noted that the accuracy of the prediction depends on the accuracy of the navigation condition prediction and the precision of the thermal inertial model. Ship navigation plans are generally highly reliable, but temporary adjustments may still occur during actual navigation. Therefore, a rolling update approach can be adopted, re-acquiring the latest navigation condition information every 30 seconds and updating the predicted temperature trajectory to cope with plan changes.

[0129] S503. Determine whether there is an area below the preset low temperature threshold in the predicted temperature trajectory. If so, start the heating device connected to the selective catalytic reduction reactor in advance at the current moment so that the catalyst bed reaches the preset target temperature range before entering the low temperature condition.

[0130] The preset low-temperature threshold refers to the critical temperature at which the catalyst activity begins to decline significantly, typically around 260 degrees Celsius for vanadium-based catalysts and around 200 degrees Celsius for zeolite-based catalysts. The target temperature range refers to the catalyst temperature range that ensures normal denitrification efficiency, typically between 280 and 450 degrees Celsius. If the predicted temperature trajectory includes a range below the low-temperature threshold, it indicates that without active heating measures, the SCR reactor will enter a low-temperature deactivation state within a certain period, potentially leading to a sharp drop in denitrification efficiency and ammonium bisulfate condensation and blockage.

[0131] In some implementations, the decision logic for active heating is as follows: First, find the lowest temperature from the predicted temperature trajectory. and the time of its occurrence Then, based on the thermal inertia model, the required heating power is determined through inverse calculation or by using optimization methods. and heating start-up time This allows the catalyst bed temperature to be raised to the target temperature range before the low-temperature range is reached. A simplified strategy is: if the future... If the catalyst temperature falls below the threshold after a few minutes, the heating device should be activated immediately and heated at maximum power until the catalyst temperature rises back above the target temperature range. A better strategy is to solve a minimum energy consumption optimization problem: The constraint is: within the predicted low temperature range And heating power This optimization problem can be solved using dynamic programming or rolling time-domain optimization.

[0132] It should be noted that the heating device can be a burner (using fuel or gas combustion to heat exhaust gas), an electric heater (using electrical energy for heating), or a heat exchanger that utilizes waste heat from the engine coolant. After starting the heating device, the catalyst bed temperature should be continuously monitored. Once the temperature reaches the target range, the heating power can be reduced or the device can be operated intermittently to maintain the temperature.

[0133] Based on the above technical solution, this application achieves pre-regulation of the thermal state of the selective catalytic reduction reactor by constructing a thermal inertia model, predicting temperature trajectories, and making active heating decisions. This active thermal state pre-regulation method makes full use of navigation condition prediction information, proactively addresses low-temperature conditions, effectively avoids catalyst deactivation at low temperatures and ammonium bisulfate condensation and blockage, and improves the operational reliability and denitrification efficiency of the ship's exhaust gas denitrification system across the entire operating range.

[0134] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0135] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A smart control method for denitrification of ship exhaust gas, applied to a ship exhaust gas aftertreatment system, the aftertreatment system comprising a selective catalytic reduction reactor and a urea injection device, characterized in that, include: An engine source nitrogen oxide predictor based on a gray box model is constructed and executed to predict the source nitrogen oxide concentration before entering the selective catalytic reduction reactor based on the acquired real-time engine operating parameters, and to generate feedforward information. A virtual ammonia coverage observer based on extended Kalman filtering is constructed and executed to estimate the ammonia coverage on the catalyst surface in the selective catalytic reduction reactor in real time based on the inlet and outlet sensor signals of the selective catalytic reduction reactor. A model predictive controller based on ammonia coverage constraint is constructed and executed. The feedforward information is used as the reference input, the ammonia coverage on the catalyst surface is used as the state feedback, and the urea injection command is obtained through online rolling optimization calculation with the constraint objective of suppressing ammonia escape. Based on the urea injection command, the urea injection device is controlled to inject urea into the exhaust gas to catalytically reduce nitrogen oxides in the exhaust gas.

2. The intelligent control method for denitrification of ship exhaust gas according to claim 1, characterized in that, The selective catalytic reduction reactor is a device for providing a site for catalytic reaction, used to enable nitrogen oxides and reducing agents to undergo a selective catalytic reduction reaction under the action of a catalyst to generate nitrogen gas and water; The real-time operating parameters include engine speed, engine load, fuel injection quantity, boost pressure, exhaust temperature, exhaust flow rate, and excess air coefficient; wherein, the excess air coefficient represents the ratio of the actual mass of air entering the internal combustion engine cylinder to the mass of air required for theoretical complete combustion.

3. The intelligent control method for denitrification of ship exhaust gas according to claim 1, characterized in that, The method further includes acquiring predicted navigation condition information and pre-adjusting the active thermal state of the selective catalytic reduction reactor based on the navigation condition information, specifically including: A thermal inertial model of the selective catalytic reduction reactor is constructed. The thermal inertial model describes the dynamic response characteristics of the catalyst bed temperature as a function of exhaust temperature and active heating power using first-order inertia plus a pure delay element. Based on the predicted navigation condition information and the thermal inertia model, the catalyst bed temperature of the selective catalytic reduction reactor is predicted within a preset time period in the future, and the predicted temperature trajectory is obtained. If there is a range in the predicted temperature trajectory that is lower than the preset low temperature threshold, the heating device connected to the selective catalytic reduction reactor will be started in advance at the current moment so that the catalyst bed reaches the preset target temperature range before entering the low temperature condition.

4. The intelligent control method for denitrification of ship exhaust gas according to claim 3, characterized in that, The method for obtaining the predicted navigation condition information is as follows: the ship's speed change trajectory, port entry and exit plan, or engine load change trajectory within a preset time period are obtained through the ship navigation planning system or the ship autopilot system, and used as the predicted navigation condition information.

5. The intelligent control method for denitrification of ship exhaust gas according to claim 1, characterized in that, The engine source nitrogen oxide predictor adopts a gray box modeling architecture, including a mechanism model part and a black box correction model part; The mechanistic model is based on the extended Zeldovich mechanism. It uses the combustion characteristic parameters in the engine's real-time operating parameters to calculate the in-cylinder characteristic temperature and calculates the nitrogen oxide generation rate based on the Arrhenius equation to obtain the mechanistic model output value. The black-box correction model part adopts a radial basis function network. Taking the real-time operating parameters of the engine as input, it outputs the deviation correction value between the output value of the mechanism model and the actual source nitrogen oxide concentration, and obtains the output value of the black-box correction model, which is used to compensate for the prediction deviation caused by combustion non-ideal in the mechanism model part. The engine source nitrogen oxide predictor obtains the predicted source nitrogen oxide concentration by adding the output value of the mechanism model to the output value of the black box correction model.

6. The intelligent control method for denitrification of ship exhaust gas according to claim 5, characterized in that, The parameters of the black-box correction model are obtained through offline identification. The offline identification aims to minimize the root mean square error between the predicted value and the measured value, and introduces a regularization term. The regularization term includes a weight decay term and a gradient constraint term. The gradient constraint term is used to force the gradient of the black-box correction model part with respect to the engine operating parameters to remain orthogonal to the gradient of the mechanism model part with respect to the engine operating parameters, so that the black-box correction model part only compensates for the nonlinear residuals not captured by the mechanism model part, without changing the main physical law of nitrogen oxide generation changing with operating conditions reflected by the mechanism model part.

7. The intelligent control method for denitrification of ship exhaust gas according to claim 1, characterized in that, The virtual ammonia coverage observer is built on an extended Kalman filter framework and includes a state prediction module, an observation update module, and a sensor signal decoupling module. The virtual ammonia coverage observer establishes a discrete state space model. The discrete state space model uses the ammonia coverage on the catalyst surface and the gas phase ammonia concentration in the selective catalytic reduction reactor as state variables, the ammonia concentration corresponding to the urea injection rate as the control input, and the exhaust temperature and exhaust flow rate as time-varying parameters to construct a state transition equation. The sensor signal decoupling module constructs an observation equation based on the cross-sensitivity characteristics of the nitrogen oxide sensor at the outlet of the selective catalytic reduction reactor to ammonia. The observation equation adopts a dual-timescale dynamic separation model, which expresses the measurement signal of the nitrogen oxide sensor as the superposition of the actual nitrogen oxide concentration response and the actual ammonia concentration response at the outlet. The actual nitrogen oxide concentration response is characterized by a first time constant, and the actual ammonia concentration response is characterized by a second time constant, and the second time constant is smaller than the first time constant. The state prediction module and the observation update module perform extended Kalman filter recursion based on the state transition equation and the observation equation, output the estimated value of ammonia coverage on the catalyst surface in real time, and output the estimated value as a state feedback signal to the model prediction controller.

8. The intelligent control method for denitrification of ship exhaust gas according to claim 1, characterized in that, The model predictive controller is built on a rolling time-domain optimization framework and includes a predictive model module, a constraint processing module, and an optimization solution module. The prediction model module establishes an ammonia storage kinetic model for predictive control. The ammonia storage kinetic model describes the dynamic change of ammonia coverage on the catalyst surface using a first-order equation of state. The ammonia concentration corresponding to the urea injection command is used as the control input, the source nitrogen oxide concentration corresponding to the feedforward information is used as the known disturbance input, and the ammonia coverage on the catalyst surface is used as the current value of the state variable. In each control cycle, the constraint processing module calculates the upper limit constraint value of the ammonia coverage on the catalyst surface based on the currently measured exhaust temperature and space velocity, and sets the control input amplitude constraint. The optimization solution module takes minimizing the sum of squares of the deviation between the predicted outlet nitrogen oxide concentration and the target in the time domain as the optimization objective, and constructs and solves the rolling time domain optimization problem with the control input amplitude constraint and the upper bound constraint as the constraint conditions. The first element of the optimal control sequence obtained by the solution is used as the urea injection command output of the previous time step.

9. The intelligent control method for denitrification of ship exhaust gas according to claim 8, characterized in that, When the constraint processing module calculates the upper limit constraint value of ammonia coverage on the catalyst surface, it is based on the adsorption reaction rate constant, desorption reaction rate constant, saturated gas phase ammonia concentration, and space velocity parameter. The upper limit constraint value increases with increasing exhaust temperature and decreases with increasing space velocity.

10. A smart control system for denitrification of ship exhaust gas, characterized in that, It includes an engine-source nitrogen oxide prediction module, a virtual ammonia coverage observation module, and a model prediction and control module; among which, The engine source nitrogen oxide prediction module is used to predict the source nitrogen oxide concentration before entering the selective catalytic reduction reactor based on the acquired real-time engine operating parameters, and generate feedforward information. The virtual ammonia coverage observation module is used to estimate the ammonia coverage on the catalyst surface in the selective catalytic reduction reactor in real time based on the inlet and outlet sensor signals of the selective catalytic reduction reactor. The model prediction and control module is used to take the feedforward information as a reference input, the ammonia coverage on the catalyst surface as a state feedback, and take suppressing ammonia escape as a constraint objective to obtain urea injection commands through online rolling optimization calculations, so as to control the urea injection device to inject urea into the exhaust gas.

Images