Adaptive data-driven modeling method for diesel selective catalytic reduction system

Abstract

The application discloses a diesel engine selective catalytic reduction system adaptive data-driven modeling method, and belongs to the field of diesel engine urea SCR emission after-treatment technology. The application aims to design a data-driven SCR model based on an adaptive echo state network, and provides an effective offline parameter identification strategy, so that a high-precision model can be directly used as a diesel engine selective catalytic reduction system adaptive data-driven modeling method facing a control model. The steps of the application are as follows: constructing an SCR data-driven model structure, and designing an adaptive parameter updating strategy. The adaptive echo state network framework used in the application has the advantages of small volume, fast operation speed, strong adaptive capacity and the like, so that the model trained according to data can be directly used as a control model.

Description

Technical Field

[0001] This invention belongs to the field of diesel engine urea SCR emission aftertreatment technology. Background Technology

[0002] Currently, the diesel engine technology widely used in my country suffers from nitrogen oxides (NOx). The issue of high emissions is a concern. The current national standard, "Limits and Measurement Methods for Pollutant Emissions from Heavy-Duty Diesel Vehicles (China Stage VI)," requires that under WHTC (World Harmonized Transient Cycle) conditions, nitrogen oxide emissions must not exceed 0.46 g / kWh. Among nitrogen oxide treatment methods, Selective Catalytic Reduction (SCR) is the most widely used and key technology.

[0003] SCR uses urea as a reducing agent to catalytically reduce the gases in diesel engine exhaust. To ensure that SCR processes most of the Under the premise of minimizing ammonia ( Traditional open-loop control and map-based control methods are increasingly inadequate for addressing escape and urea consumption issues, while model-based control methods are becoming more sophisticated, offering excellent performance and promising prospects. Designing and developing a control-oriented SCR model is fundamental to realizing model-based control methods.

[0004] Traditional mechanism-based SCR modeling methods cannot well fit the aging problem of SCR; the longer the system runs, the more obvious the error accumulation becomes. Meanwhile... The rear sensors suffer from cross-sensitivity, preventing the system from using feedback information to correct the mechanistic model in real time, thus hindering error correction. Furthermore, different engines use different types of catalysts and have varying catalyst structures, making it difficult to capture their physical properties. These factors contribute to poor portability and durability of the mechanistic model, high research costs, and hindering industrial applications.

[0005] In contrast, data-driven models, after learning from a large amount of experience, can accurately capture the cross-sensitivity characteristics under different operating conditions, and can utilize this information... Sensor feedback information enables closed-loop correction of the system, thus possessing high durability, stability, and portability, and represents the future direction of development. In order to conveniently and quickly establish high-precision data-driven SCR models in practical engineering projects, the following problems and challenges must be addressed.

[0006] The main challenges and issues include: 1. How to design the model structure to maintain accuracy while reducing size and computational complexity; 2. Which parameters need to be identified offline in the data-driven model, and how to identify them; 3. How to design online model training rules and effectively utilize the results. Real-time feedback information from the sensors enables closed-loop correction of the model. Summary of the Invention

[0007] The purpose of this invention is to design a data-driven SCR model based on an adaptive echo state network and to provide an effective offline parameter identification strategy. The resulting high-precision model can be directly used as an adaptive data-driven modeling method for diesel engine selective catalytic reduction systems oriented towards control models.

[0008] The steps of this invention are:

[0009] S1. Constructing the SCR data-driven model structure

[0010] The input to the data-driven model is defined as follows:

[0011] (1)

[0012] in, For data-driven models The input vector at time t, for Constant intake air temperature, for Constant exhaust mass flow rate for At any given time, the inlet concentration of nitrogen oxides for Urea injection rate at all times for After a moment Sensor readings;

[0013] The output and expected value of the data-driven model are as follows:

[0014] (2)

[0015] The state update equation and output equation of the SCR data-driven model in the adaptive echo state network framework are as follows:

[0016] (3)

[0017] (4)

[0018] in for The state vector of the time-adaptive echo state network reservoir. , , For state memory factors, The input weight matrix size factor, For the feedback weight matrix size factor, The spectral radius of the reservoir matrix, For the input weight matrix, For the reserve pool weight matrix, For the feedback weight matrix, To output the weight matrix;

[0019] Network activation function Using a modified hyperbolic tangent function:

[0020] (5)

[0021] in As input to the activation function, , The deformation factor is the parameter to be identified offline.

[0022] Due to the input weight matrix of the network Feedback weight matrix and output weight matrix All values ​​are randomly generated in the range [-1, 1]; the reserve pool weight matrix The following rules shall be adopted:

[0023] a) Initialize the parameters to be identified offline: connection density constant Distance constant ;

[0024] b) When a new neuron is generated, its connection probability with existing neurons is calculated using the following formula:

[0025] (6)

[0026] in Let be the Euclidean distance between the two neural nodes. A random number between [0, 1];

[0027] c) Generate a random number, if it is less than Then a connection weight is generated between [-1, 1]. , ;

[0028] d) Repeat the above two steps until the total number of neurons reaches 100, then complete the generation of the reserve pool weight matrix;

[0029] S2. Design an adaptive parameter update strategy

[0030] The root mean square error of the system at time step is:

[0031] (7)

[0032] but At time t, its derivative with respect to one of the above parameters is:

[0033] (8)

[0034] in for Error matrix at time step For a certain online identification parameter, The derivative of the input matrix with respect to this parameter is a zero vector;

[0035] make From equations (3) and (5), we have:

[0036] (9)

[0037] (10)

[0038] (11)

[0039] (12)

[0040] (13)

[0041] (14)

[0042] The final online update parameter expression is determined as follows:

[0043] (15)

[0044] in For parameters The online learning rate is a parameter to be identified offline;

[0045] Will Decomposed into and Two parts, making , and then , For any two sets of reserve pool states and have:

[0046] (16)

[0047] like Then when hour ;

[0048] For online identification of output weight matrix Let the Lyapunov function be:

[0049] (17)

[0050] The difference is then divided into:

[0051] (18)

[0052] Design the output weight matrix Update rules:

[0053] (19)

[0054] in Substituting into equation (16), we get:

[0055] (20)

[0056] When constant When using this strategy to update the output weight matrix, it can be guaranteed that... hour, .

[0057] The online parameter update strategy and offline parameter identification method of the model of this invention:

[0058] a) Set the optimization period and population size, set the upper and lower limits of each parameter, set position A to the initial value given in Table 2, and set positions B, C and other particles to random values ​​between the upper and lower limits;

[0059] b) Calculate the fitness of all particles, sort them from low to high fitness, and assign the top three fitness values ​​to A, B, and C respectively. The fitness value is the mean squared error of the SCR model trained using data with certain particle parameters on the test set.

[0060] c) Update the particle position using the following formula:

[0061] (twenty one)

[0062] in The set number of optimization cycles, , , , , , All numbers are random numbers between [0, 2]. For the first The particle in the first The period is located in the parameter space;

[0063] d) Repeat the above two steps until... Stop seeking optimization.

[0064] This invention addresses the selective catalytic reduction (SCR) system of diesel engines, proposing an adaptive echo state network (ESN) modeling method and a corresponding offline parameter identification strategy. The ESN is a type of neural network with memory capabilities, capable of extracting state-space-like equations, which facilitates controller design compared to other data-driven models. The adaptive ESN framework used in this invention offers advantages such as small size, fast operation, and strong adaptability; the model trained using this framework can be directly used as a control-oriented model. Attached Figure Description

[0065] Figure 1 Diagram of the adaptive echo state network structure;

[0066] Figure 2 After data-driven SCR model Sensor verification comparison chart;

[0067] Figure 3 The traditional echo state network data-driven SCR model and the adaptive echo state network data-driven SCR model are then compared. Sensor error comparison chart. Detailed Implementation

[0068] To improve the accuracy of diesel engine SCR exhaust aftertreatment models, this invention designs a data-driven SCR model based on an adaptive echo state network and provides an effective offline parameter identification strategy. The resulting high-precision model can be directly used as a control-oriented model.

[0069] The objective of this invention is achieved through the following technical solution: adaptive data-driven SCR model establishment and parameter identification, including the following steps:

[0070] Step 1: Fully stimulate the controlled object (China VI diesel engine test bench), collect input and output data, and divide the data into training set and test set;

[0071] Step 2: Based on the adaptive echo state network framework, construct the basic structure of the SCR data-driven model;

[0072] Step 3: Design an adaptive parameter update strategy based on gradient descent and an output weight update strategy based on Lyapunov stability proof;

[0073] Step 4: Based on the complexity of the system's input and output data, select the offline parameters to be identified and use an intelligent optimization algorithm for identification;

[0074] Step 5: Verify the actual effect of the data-driven SCR model on a simulation platform or a real test bench.

[0075] The engine experiments related to this invention were conducted on a test bench meeting the China VI emission standard, and the information collected included the inlet of the SCR exhaust aftertreatment system. Concentration, inlet urea injection rate, inlet air temperature, inlet air flow rate, outlet... Information from sensors and other sources is obtained from the original sensors.

[0076] This invention targets diesel engines using SCR exhaust aftertreatment technology. To address the problem that the SCR mechanism model has poor portability and durability and is difficult to meet practical needs, this invention proposes an SCR data-driven modeling method based on an adaptive echo state network.

[0077] This invention employs the Grey Wolf algorithm to identify some parameters of the model offline, thereby fully stimulating the potential of the data-driven model and making it more suitable for selective catalytic reduction systems.

[0078] The implementation of this invention specifically includes the following steps:

[0079] 1. Incentivize the controlled object to acquire data.

[0080] Perform multiple sets of constant intake temperature, constant Experiments were conducted at inlet concentration, constant exhaust mass flow rate, and constant urea injection rate to investigate the ammonia storage characteristics of the SCR system. This data was then used as training data for a data-driven model. Experimental data under WHTC operating conditions were used as the test and validation sets for the data-driven model. Before training or validating the model, the input data needed to be normalized to prevent multicollinearity caused by activation function saturation, which could lead to information loss.

[0081] 2. Construct the SCR data-driven model structure

[0082] First, determine the input to the data-driven model as follows:

[0083] (1)

[0084] in, For data-driven models The input vector at time t, for Constant intake air temperature, for Constant exhaust mass flow rate for At any given time, the inlet concentration of nitrogen oxides for Urea injection rate at all times for After a moment Sensor readings.

[0085] The output and expected value of the data-driven model are as follows:

[0086] (2)

[0087] The state update equation and output equation of the SCR data-driven model based on the adaptive echo state network framework are as follows:

[0088] (3)

[0089] (4)

[0090] in for The state vector of the time-adaptive echo state network reservoir, and the meaning and update method of the other parameters and matrices are shown in Table 1.

[0091] Table 1: Meaning of parameters and matrices in the state update equation and output equation

[0092]

[0093] Network structure such as Figure 1 As shown.

[0094] Network activation function Using a modified hyperbolic tangent function:

[0095] (5)

[0096] in As input to the activation function, , The deformation factor is the parameter to be identified offline.

[0097] Due to the input weight matrix of the network Feedback weight matrix and output weight matrix All values ​​are randomly generated in the range [-1, 1]. Reserve pool weight matrix. The following rules shall be adopted:

[0098] a) Initialize the parameters to be identified offline: connection density constant Distance constant ;

[0099] b) When a new neuron is generated, its connection probability with existing neurons is calculated using the following formula:

[0100] (6)

[0101] in Let be the Euclidean distance between the two neural nodes. A random number between [0, 1];

[0102] c) Generate a random number, if it is less than Then a connection weight is generated between [-1, 1]. , ;

[0103] d) Repeat the above two steps until the total number of neurons reaches 100, then complete the generation of the reserve pool weight matrix.

[0104] 3. Design an adaptive parameter update strategy

[0105] In summary, the system has 6 online identification parameters and 1 online identification matrix. For the online identification parameters, a gradient descent-like update strategy is adopted.

[0106] The root mean square error of the system at time step is:

[0107] (7)

[0108] but At time t, its derivative with respect to one of the above parameters is:

[0109] (8)

[0110] in for Error matrix at time step For a certain online identification parameter, The derivative of the input matrix with respect to this parameter is a zero vector.

[0111] make From equations (3) and (5), we have:

[0112] (9)

[0113] (10)

[0114] (11)

[0115] (12)

[0116] (13)

[0117] (14).

[0118] The final online update parameter expression is determined as follows:

[0119] (15)

[0120] in For parameters The online learning rate is the parameter to be identified offline. In this process, only... This allows the system to maintain its echo state characteristic, meaning that the output of the neural nodes in the reservoir is ultimately only related to the external input.

[0121] Will Decomposed into and Two parts, making , and then , For any two sets of reserve pool states and have:

[0122] (16)

[0123] like Then when hour .

[0124] For online identification of output weight matrix Let the Lyapunov function be:

[0125] (17)

[0126] The difference is then divided into:

[0127] (18).

[0128] Based on this design, the output weight matrix is... Update rules:

[0129] (19)

[0130] in Substitute into equation (16).

[0131] We can obtain:

[0132] (20)

[0133] When constant When using this strategy to update the output weight matrix, it can be guaranteed that... hour, .

[0134] 4. Use intelligent optimization algorithms to identify some parameters offline.

[0135] By updating some parameters online, the data-driven SCR model based on the adaptive echo state network has achieved high portability and stability. In order to specialize its characteristics on different engines and improve the model accuracy, it is also necessary to perform offline identification of some parameters based on the posterior.

[0136] In summary, the parameters to be identified offline are: , , , , , The online parameter learning rate There are six parameters in total, so there are a total of 11 offline identification parameters.

[0137] This invention uses an intelligent optimization algorithm to accomplish this task, taking the Grey Wolf algorithm as an example. The Grey Wolf algorithm is an intelligent optimization algorithm that searches the parameter space by simulating the hunting activities of a grey wolf to find the global optimum. Due to its strong convergence performance, it can achieve good optimization results with the given favorable initial conditions, making it suitable for offline parameter identification of data-driven SCR models. The initial parameter values ​​based on the Grey Wolf algorithm are shown in Table 2.

[0138] Table 2: Initial values ​​of offline identification parameters for the data-driven SCR model based on the Grey Wolf algorithm

[0139]

[0140] The algorithm process can be divided into the following four steps:

[0141] a) Set the optimization period and population size, set the upper and lower limits of each parameter, set position A to the initial value given in Table 2, and set positions B, C and other particles to random values ​​between the upper and lower limits;

[0142] b) Calculate the fitness of all particles, sort them from low to high fitness, and assign the top three fitness values ​​to A, B, and C respectively. The fitness value is the mean squared error of the SCR model trained using data with certain particle parameters on the test set.

[0143] c) Update the particle position using the following formula:

[0144] (twenty one)

[0145] in The set number of optimization cycles, , , , , , All numbers are random numbers between [0, 2]. For the first The particle in the first The period is located in the parameter space;

[0146] d) Repeat the above two steps until... Stop seeking optimization.

[0147] Verify model accuracy

[0148] After offline parameter identification, the final effect of the model should be viewed using a simulation model or a real test bench. This invention... The sensor's performance was verified under WHTC operating conditions as follows: Figure 2 As shown, Figure 3 This is a comparison chart of the errors between the traditional echo state network data-driven SCR model and the adaptive echo state network data-driven SCR model.

[0149] The data-driven SCR model and its offline parameter identification method proposed in this invention have the following advantages:

[0150] 1) The data-driven SCR model based on the adaptive echo state network has high portability and durability while ensuring high accuracy.

[0151] 2) Offline parameter identification within the initial parameter range given in this invention enables the data-driven model to quickly capture the characteristics of SCR systems with different structures.

[0152] 3) The data-driven model proposed in this invention is small in size and fast in speed, making it more suitable for practical engineering applications compared to other data-driven models.

[0153] This invention proposes a data-driven model for SCR based on an adaptive echo state network, along with a quick method for constructing the various structures of the model. Based on stability, online parameter updates and offline parameter identification methods for the data-driven SCR model are presented. Initial values ​​for the offline identification parameters of the engine SCR system, which have high general applicability, are provided, enabling the data-driven model to quickly capture the characteristics of SCR systems with different structures.

Claims

1. An adaptive data-driven modeling method for a diesel engine selective catalytic reduction system, characterized in that: The steps are as follows: S1. Constructing the SCR data-driven model structure The input to the data-driven model is defined as follows: （1） in, For data-driven models The input vector at time t, for Constant intake air temperature, for Constant exhaust mass flow rate for At any given time, the inlet concentration of nitrogen oxides for Urea injection rate at all times for After a moment Sensor readings; The output and expected value of the data-driven model are as follows: （2） The state update equation and output equation of the SCR data-driven model in the adaptive echo state network framework are as follows: （3） （4） in for The state vector of the time-adaptive echo state network reservoir. , , For state memory factors, The input weight matrix size factor, For the feedback weight matrix size factor, The spectral radius of the reservoir matrix, For the input weight matrix, For the reserve pool weight matrix, For the feedback weight matrix, To output the weight matrix; Network activation function Using a modified hyperbolic tangent function: （5） in As input to the activation function, , The deformation factor is the parameter to be identified offline. Due to the input weight matrix of the network Feedback weight matrix and output weight matrix All values ​​are randomly generated in the range [-1, 1]; the reserve pool weight matrix The following rules shall be adopted: a) Initialize the parameters to be identified offline: connection density constant Distance constant ; b) When a new neuron is generated, its connection probability with existing neurons is calculated using the following formula: （6） in Let be the Euclidean distance between the two neural nodes. A random number between [0, 1]; c) Generate a random number, if it is less than Then a connection weight is generated between [-1, 1]. , ; d) Repeat steps b) and c) until the total number of neurons reaches 100, then complete the generation of the reserve pool weight matrix; S2. Design an adaptive parameter update strategy The root mean square error of the system at time step is: （7） but At time t, the derivatives of the state update equation and the output equation with respect to the parameters are: （8） in for Error matrix at time step For a certain online identification parameter, The derivative of the input matrix with respect to this parameter is a zero vector; make From equations (3) and (5), we have: （9） （10） （11） （12） （13） （14） The final online update parameter expression is determined as follows: （15） in For parameters The online learning rate is a parameter to be identified offline; Will Decomposed into and Two parts, making , and then , For any two sets of reserve pool states and have: （16） like Then when hour ; For online identification of output weight matrix Let the Lyapunov function be: （17） The difference is then divided into: （18） Design the output weight matrix Update rules: （19） in Substituting into equation (16), we get: （20） When constant When using this strategy to update the output weight matrix, it can be guaranteed that... hour, .

2. The adaptive data-driven modeling method for diesel engine selective catalytic reduction systems according to claim 1, characterized in that: Online parameter update strategy and offline parameter identification method for the model: a) Set the optimization cycle and population size, set the upper and lower limits of the offline identification parameters of the data-driven SCR model of the Grey Wolf algorithm, set the particle at position A as the initial value, and set all particles except the particle at position A as random values ​​between the upper and lower limits, where all particles include the particles at positions B and C. b) Calculate the fitness of all particles containing positions A, B, and C, sort them from low to high fitness, and assign the top three values ​​to A, B, and C respectively; the fitness value is the mean squared error of the SCR model driven by data with a certain particle parameter on the test set after training. c) Update the particle position using the following formula: （21） in The set number of optimization cycles, , , , , , All numbers are random numbers between [0, 2]. For the first The particle in the first The period is located in the parameter space; d) Repeat steps b) and c) until... Stop seeking optimization.

Images