Method for predicting exhaust NOx concentration of diesel engine SCR system

Abstract

The method for predicting tail emission NOx concentration of diesel engine SCR system belongs to the field of engine emission control. The method solves the problems that the existing tail emission of SCR is not predicted in the aftertreatment development, and in the existing tail gas emission prediction method, the hyperparameters of the model are determined by pre-experiment, and the process of optimizing the hyperparameters of the prediction model is lacked, resulting in low accuracy of the prediction result. The method comprises the following steps: firstly, constructing a data set, determining that each sample in the data set is composed of 8 parameters, and adding a concentration output label to each sample; secondly, building an SCR system concentration prediction model by using an LSTM neural network; thirdly, determining the hyperparameters to be optimized in the SCR system concentration prediction model, and performing optimization to obtain the optimal SCR system concentration prediction model; and finally, predicting the tail emission NOx concentration of the SCR system located downstream of the diesel engine by using the model. The method is mainly used for predicting the tail emission NOx concentration emitted into the air.

Description

Technical Field

[0001] This invention belongs to the field of engine emission control. Background Technology

[0002] With the continuous development of diesel engine technology and increasingly stringent emission requirements, higher demands are being placed on after-treatment systems. Different emission control technologies have emerged during the development of after-treatment systems. SCR (Sequencing-Reduction) will be the main technological direction for future diesel engine emission upgrades in China, further treating diesel engine exhaust emissions. Specifically, SCR technology reduces NOx emissions by precisely controlling the amount of urea injected, thereby lowering the concentration of NOx released into the air. Malfunctions in the SCR system will directly lead to increased NOx emissions, causing environmental pollution. Therefore, it is crucial to ensure that NOx emissions meet limits after SCR treatment.

[0003] Diesel engines operate under complex conditions, and directly measuring NOx concentration in exhaust emissions is affected by sensor accuracy and reliability, and suffers from time lag and data distortion. To achieve accurate NOx measurement, two approaches can be taken: firstly, developing next-generation sensors to collect data that approximates the true value as closely as possible; secondly, establishing deep learning models based on big data from vehicle operations to accurately predict diesel engine exhaust NOx concentration in real time. The former approach involves sensor construction and manufacturing principles, relying on advancements in design technology, and is a slow process. However, with the development of vehicle networking and neural networks, it has become possible to monitor NOx concentration in real time by establishing high-precision NOx prediction models.

[0004] There are many methods for predicting NOx emissions through exhaust NOx prediction models, but most of them are for treating the exhaust emissions of diesel engines, rather than predicting the NOx emitted by SCR during the development of diesel engine aftertreatment. Among them, the existing patent documents related to the direct exhaust emission prediction method of the engine are as follows: (1) Yu Shu et al. applied for a vehicle exhaust emission prediction method and system based on machine learning algorithm, application number: 202011041577.9; (2) Zhang Yujun et al. applied for a road motor vehicle exhaust emission prediction method based on improved attention bidirectional long short-term memory network, application number: 202010443547.4; (3) Ling Qiang et al. applied for a vehicle exhaust emission prediction method and system based on multi-component fusion time network, application number: 202011507315.7.

[0005] In the above inventions, patent (1) uses a convolutional neural network to predict NOx in the exhaust; patent (2) uses a training set to determine the hyperparameters of the model through pre-experimentation, and uses a single variable to adjust the hyperparameters to obtain a relatively optimal combination of multiple hyperparameters. The entire tuning process is complex and time-consuming; patent (3) uses a one-dimensional convolutional neural network to predict NOx in the exhaust. All of the above are predictions of exhaust directly emitted by diesel engines. There is a lack of a method for predicting exhaust after processing by the SCR system in the after-treatment development. In the existing exhaust emission prediction methods, the hyperparameters of the model are determined through pre-experimentation, but there is a lack of a process for optimizing the hyperparameters of the prediction model, resulting in low accuracy of the prediction results. The above problems need to be solved urgently. Summary of the Invention

[0006] The purpose of this invention is to address the problems of low accuracy in existing exhaust emission prediction methods, which rely on pre-experimental determination of model hyperparameters and lack optimization of these hyperparameters. This invention provides a method for predicting NOx concentration in exhaust emissions from diesel engine SCR aftertreatment systems.

[0007] A method for predicting NOx concentration in diesel engine SCR systems, comprising the following steps:

[0008] Step 1: Construct a dataset for NOx prediction based on SCR system data collected under real historical operating conditions; and add a concentration output label to each sample in the dataset.

[0009] Each sample consists of 8 parameters: diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, average SCR ammonia storage ratio, average urea pressure, reducing agent injection quantity, and exhaust flow rate.

[0010] The concentration output label for each sample is the NOx concentration output value of the SCR system at the corresponding time point;

[0011] Step 2: Build an SCR system concentration prediction model using an LSTM neural network;

[0012] Step 3: Determine the hyperparameters to be optimized in the SCR system concentration prediction model, and construct a population matrix X in the search space using the hyperparameters to be optimized; wherein, the population matrix X includes N sets of hyperparameter combinations, and the hyperparameters in each set of hyperparameter combinations have the same type composition; each set of hyperparameter combinations includes d hyperparameters, and each hyperparameter corresponds to a type;

[0013] Then, the constructed sample set is used to train the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X. The optimal hyperparameter combination is searched from the population matrix X, and the SCR system concentration prediction model under the optimal hyperparameter combination is taken as the optimal SCR system concentration prediction model.

[0014] Step 4: Use the current diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, SCR ammonia average storage ratio, urea pressure average value, reductant injection demand, and exhaust flow rate as inputs to the optimal SCR system concentration prediction model. Perform concentration prediction through the optimal SCR system concentration prediction model and output the current true value of NOx concentration in the SCR system exhaust.

[0015] Preferably, in step one, the specific process of constructing the dataset used for NOx prediction includes:

[0016] Step 1: Collect various types of parameters of the SCR system under real historical operating conditions within a preset time period;

[0017] Step 1 and Step 2: Filter the various types of parameters of the SCR system collected at each time point, and select the 8 parameters corresponding to each time point, namely: diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, SCR ammonia average storage ratio, urea pressure average value, reducing agent injection quantity and exhaust flow rate.

[0018] Step 13: Preprocess the 8 parameters corresponding to each time point to remove null values ​​and outliers, and normalize the 8 parameters corresponding to each time point after preprocessing. Then, use the 8 parameters corresponding to each time point after normalization as a sample.

[0019] Preferably, in step three, the population matrix...

[0020] X i This represents the position of the i-th hyperparameter combination in the search space, where i = 1, 2, ..., N; x ij Let represent the position of the i-th hyperparameter combination in the j-th dimension, where j = 1, 2, ..., d.

[0021] Preferably, in step three, the specific process of training the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X using the constructed sample set, and searching for the optimal hyperparameter combination from the population matrix X, includes:

[0022] In step three, set the iteration threshold H and the threshold of each hyperparameter to be optimized, and initialize the population matrix X within the search space formed by the thresholds of each hyperparameter to be optimized.

[0023] In step 3.2, the initial value of h is 1. During the h-th iteration, the constructed sample set is used to train the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X, thereby obtaining the fitness function vector of the population matrix X.

[0024] Where, f(x) i,1 x i,2 , ..., x i,d f is the fitness function corresponding to the i-th hyperparameter combination; i = 1, 2, ..., N; f i f(x) i,1 x i,2 , ..., x i,d The value of );

[0025] Step 3: In the h-th iteration, the combination of hyperparameters corresponding to the minimum fitness function value in the fitness function vector F is taken as the initial optimal hyperparameter combination.

[0026] Steps 3 and 4: Determine whether h has reached the iteration threshold H. If yes, proceed to step 3 and 6; if no, proceed to step 3 and 5.

[0027] Step 35: Update the population matrix X in the search space, let h = h + 1, and execute step 32;

[0028] Step 36: Select the minimum value of the fitness function value among all the initial optimal hyperparameter combinations obtained under all iterations, and take the initial optimal hyperparameter combination corresponding to the minimum fitness function value as the optimal hyperparameter combination.

[0029] Preferably, in step three-five, when updating the population matrix X within the search space, x in the population matrix X i,j The update methods include:

[0030]

[0031] Where, x best The optimal position in the j-th dimension at the current iteration number is the fitness function value corresponding to this optimal position, which is the minimum value in the fitness function vector F; x worst This represents the worst position in the j-th dimension at the current iteration number, and the fitness function value corresponding to this worst position is the maximum value in the fitness function vector F; β is a random number generated according to a normal distribution with a mean of 0 and a variance of 1; f wf is the maximum value of the fitness function in the fitness function vector F at the current iteration number; g ε is the minimum value of the fitness function in the fitness function vector F at the current iteration number, where ε is the coefficient and K is a random number between [-1, 1].

[0032] Preferably, in step three, the search space is composed of the threshold values ​​of each hyperparameter.

[0033] Preferably, when d is 4, the four hyperparameters in each hyperparameter combination are the number of training iterations, the learning rate, and the number of neurons in the first and second hidden layers of the LSTM neural network.

[0034] Preferably, the threshold range for the number of training iterations is [10, 100]; the threshold range for the learning rate is [0.001, 0.01]; the threshold range for the number of neurons in the first hidden layer of the neural network is [1, 100]; and the threshold range for the number of neurons in the second hidden layer of the neural network is [1, 100].

[0035] Preferably, in step three, the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X is trained using the constructed sample set, and the optimal hyperparameter combination is searched from the population matrix X and implemented using the SSA algorithm.

[0036] Preferably, in step three, the number of hyperparameters to be optimized is the same as the dimension of the search space.

[0037] This invention brings beneficial effects;

[0038] This invention constructs a dataset based on data collected under real historical operating conditions, ensuring the authenticity of the training data. Furthermore, it optimizes the hyperparameters of an SCR system concentration prediction model constructed using a Long Short-Term Memory (LSTM) neural network to obtain the optimal SCR system concentration prediction model. This model is then used to predict the NOx concentration emitted by the downstream SCR system of a real-time diesel engine, enabling real-time monitoring of NOx levels in the incoming air. This allows for adaptive adjustment of urea injection and reduces diesel engine costs, significantly contributing to energy conservation and emission reduction. The invention achieves online cloud-based monitoring of NOx levels in the air, and to avoid local optima during the hyperparameter optimization process, the optimal hyperparameter combination is obtained by continuously updating the population matrix.

[0039] This invention takes into account the parameters used in the sample data used to construct NOx prediction. When screening parameters, it comprehensively considers the parameters that affect NOx conversion efficiency and correlation analysis, thereby determining the eight parameters included in the sample.

[0040] The method for predicting NOx concentration in diesel engine SCR systems, as verified by this invention, shows a 15.9% and 1.17% reduction in the root mean square error (RMSE) of its evaluation model accuracy compared to existing MLP and LSTM models. This addresses the low prediction accuracy issue inherent in existing technologies that rely on pre-experimental determination of model hyperparameters. Furthermore, the model constructed in this invention is an SCR system concentration prediction model, predicting NOx concentration directly emitted into the air, rather than predicting NOx concentration emitted from the upstream diesel engine. This invention more accurately reflects the performance indicators of the SCR system. The SCR system can adaptively adjust urea injection and reduce diesel engine costs, which is of great significance for energy conservation and emission reduction.

[0041] The SCR system concentration prediction model of this invention has been verified to have universal applicability. The true NOx concentration in the tail-end emissions of the SCR system is collected by sensors; the SCR system concentration prediction model of this invention can fit the true sensor measurements very well. The established model is deployed in the cloud, and the model can replace sensors to realize adaptive adjustment of urea injection and emission monitoring functions; at the same time, simplifying the sensors in the aftertreatment system can improve economic efficiency. Attached Figure Description

[0042] Figure 1 This is a schematic diagram illustrating the principle of predicting NOx concentration in the exhaust of a diesel engine using an SCR system.

[0043] Figure 2 It is a graph showing the change in fitness values ​​corresponding to the initial optimal hyperparameter combination during the iteration process;

[0044] Figure 3 It is the curve showing the change in the learning rate during the optimization process;

[0045] Figure 4 It is the model training iterations optimization curve;

[0046] Figure 5 This is the curve for optimizing the number of neurons in the first hidden layer;

[0047] Figure 6 The curve for optimizing the number of neurons in the second hidden layer.

[0048] Figure 7 This is a comparison chart between the predicted value of NOx concentration in downstream SCR system emissions using the SCR system concentration prediction model of this invention and the actual value of NOx concentration in downstream SCR system emissions collected in real time.

[0049] Figure 8 This is a comparison chart of the distribution fitting between the predicted and actual NOx emission values ​​using the SCR system concentration prediction model of this invention;

[0050] Figure 9 This is a comparison chart of predicted and actual NOx values ​​for different engine models; among them, Figure 9 a is a comparison chart of the predicted and actual values ​​of engine 56718; Figure 9 b is a comparison chart of the predicted and actual values ​​of engine 66439. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0053] Example 1:

[0054] See Figure 1 This embodiment 1 describes a method for predicting NOx concentration in the exhaust of a diesel engine SCR system. The method includes the following steps:

[0055] Step 1: Construct a dataset for NOx prediction based on SCR system data collected under real historical operating conditions; and add a concentration output label to each sample in the dataset.

[0056] Each sample consists of 8 parameters: diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, average SCR ammonia storage ratio, average urea pressure, reducing agent injection quantity, and exhaust flow rate.

[0057] The concentration output label for each sample is the NOx concentration output value of the SCR system at the corresponding time point;

[0058] Step 2: Build an SCR system concentration prediction model using an LSTM neural network;

[0059] Step 3: Determine the hyperparameters to be optimized in the SCR system concentration prediction model, and construct a population matrix X in the search space using the hyperparameters to be optimized; wherein, the population matrix X includes N sets of hyperparameter combinations, and the hyperparameters in each set of hyperparameter combinations have the same type composition; each set of hyperparameter combinations includes d hyperparameters, and each hyperparameter corresponds to a type;

[0060] Then, the constructed sample set is used to train the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X. The optimal hyperparameter combination is searched from the population matrix X, and the SCR system concentration prediction model under the optimal hyperparameter combination is taken as the optimal SCR system concentration prediction model.

[0061] Step 4: Use the current diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, SCR ammonia average storage ratio, urea pressure average value, reductant injection demand, and exhaust flow rate as inputs to the optimal SCR system concentration prediction model. Perform concentration prediction through the optimal SCR system concentration prediction model and output the current true value of NOx concentration in the SCR system exhaust.

[0062] In this embodiment, each hyperparameter in the combination of hyperparameters of the population matrix X has a specific value. This invention optimizes the hyperparameters of the SCR system concentration prediction model constructed by the Long Short-Term Memory Neural Network (LSTM Neural Network) to predict the real-time NOx emissions of the actual vehicle, so as to monitor the NOx content in the air in real time. It can adaptively adjust the urea injection in the SCR system and reduce the cost of diesel engines, which is of great significance for energy conservation and emission reduction.

[0063] An SCR system concentration prediction model was built using an LSTM neural network, which was implemented using existing technology.

[0064] Furthermore, in step one, the specific process of constructing the dataset used for NOx prediction includes:

[0065] Step 1: Collect various types of parameters of the SCR system under real historical operating conditions within a preset time period;

[0066] Step 1 and Step 2: Filter the various types of parameters of the SCR system collected at each time point, and select the 8 parameters corresponding to each time point, namely: diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, SCR ammonia average storage ratio, urea pressure average value, reducing agent injection quantity and exhaust flow rate.

[0067] Step 13: Preprocess the 8 parameters corresponding to each time point to remove null values ​​and outliers, and normalize the 8 parameters corresponding to each time point after normalization. Then, use the 8 parameters corresponding to each time point after normalization as a sample.

[0068] In this preferred embodiment, the collected data is filtered and preprocessed to obtain usable data for NOx prediction. When filtering parameters, the parameters affecting NOx conversion efficiency and correlation analysis are comprehensively considered to determine the eight parameters included in the sample. At the same time, in order to eliminate the influence of different units on the data and improve the convergence speed and prediction accuracy of the model, the data is normalized to the range of [0, 1].

[0069] Furthermore, in step three, the population matrix...

[0070] X i This represents the position of the i-th hyperparameter combination in the search space, where i = 1, 2, ..., N; x ij This represents the position of the i-th hyperparameter combination in the j-th dimension, where j = 1, 2, ..., d. In practical applications, the number of hyperparameters to be optimized can be determined according to the actual application, and each type of hyperparameter represents a dimension.

[0071] Furthermore, in step three, the specific process of training the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X using the constructed sample set, and searching for the optimal hyperparameter combination from the population matrix X, includes:

[0072] In step three, set the iteration threshold H and the threshold of each hyperparameter to be optimized, and initialize the population matrix X within the search space formed by the thresholds of each hyperparameter to be optimized.

[0073] In step 3.2, the initial value of h is 1. During the h-th iteration, the constructed sample set is used to train the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X, thereby obtaining the fitness function vector of the population matrix X.

[0074] Where, f(x) i,1 x i,2 , ..., x i,d f is the fitness function corresponding to the i-th hyperparameter combination; i = 1, 2, ..., N; f i f(x) i,1 x i,2 , ..., x i,d The value of );

[0075] Step 3: In the h-th iteration, the fitness function vector F... N The combination of hyperparameters corresponding to the minimum fitness function value is taken as the initial optimal hyperparameter combination;

[0076] Steps 3 and 4: Determine whether h has reached the iteration threshold H. If yes, proceed to step 3 and 6; if no, proceed to step 3 and 5.

[0077] Step 35: Update the population matrix X in the search space, let h = h + 1, and execute step 32;

[0078] Step 36: Select the minimum value of the fitness function value among all the initial optimal hyperparameter combinations obtained under all iterations, and take the initial optimal hyperparameter combination corresponding to the minimum fitness function value as the optimal hyperparameter combination.

[0079] In this preferred embodiment, since a smaller population size makes it easier to get trapped in local optima, the selected population matrix X is continuously updated within the search space during the design process. A new population matrix X is generated for each iteration number, and the minimum value of the fitness function at the current iteration number is taken as the initial optimal hyperparameter combination. Then, the minimum value of the fitness function at all iterations is selected to avoid local optima and achieve overall optimization. The smaller the fitness function value, the better the corresponding performance.

[0080] When applying the model, the dataset can be divided into a training set and a test set. The training set is used to train the SCR system concentration prediction model, and the test set is used to test the SCR system concentration prediction model. The fitness function is the fitness function that determines the location. The mean squared error (MSE) can be selected as the fitness function.

[0081] The hyperparameters in the optimal hyperparameter combination are the best hyperparameters of the SCR system concentration prediction model. By assigning the optimal hyperparameter combination to the SCR system concentration prediction model, the optimal SCR system concentration prediction model can be obtained.

[0082] Furthermore, in step three-five, when updating the population matrix X within the search space, x in the population matrix X... i,j The update methods include:

[0083]

[0084] Where, x best The optimal position in the j-th dimension at the current iteration number is the fitness function value corresponding to this optimal position, which is the minimum value in the fitness function vector F; x worst This represents the worst position in the j-th dimension at the current iteration number, and the fitness function value corresponding to this worst position is the maximum value in the fitness function vector F; β is a random number generated according to a normal distribution with a mean of 0 and a variance of 1; f w f is the maximum value of the fitness function in the fitness function vector F at the current iteration number; gε is the minimum value of the fitness function in the fitness function vector F at the current iteration number, where ε is the coefficient and K is a random number between [-1, 1].

[0085] In this preferred embodiment, ε is a very small number to avoid the inability to update the position when the denominator is 0. K is a random number between [-1, 1]. Positive values ​​are far from the worst hyperparameter combination, and negative values ​​are close to the worst hyperparameter combination. When f i ≠f g When f is at its optimal value, it indicates that the hyperparameter combination is not at the optimal point, which is dangerous, and we should move closer to the optimal hyperparameter combination; when f i =f g When the value is zero, it indicates that the hyperparameter combination is at its optimal position for the current iteration number, and its range needs to be expanded to obtain the global optimum.

[0086] Furthermore, in step three, the search space is composed of the thresholds of each hyperparameter. The number of hyperparameters to be optimized is the same as the dimension of the search space, which is also the optimization dimension.

[0087] In practical applications, when d is 4, the four hyperparameters in each hyperparameter combination are the number of training iterations, the learning rate, and the number of neurons in the first and second hidden layers of the LSTM neural network. The threshold range for the number of training iterations is [10, 100], the threshold range for the learning rate is [0.001, 0.01], the threshold range for the number of neurons in the first hidden layer of the neural network is [1, 100], and the threshold range for the number of neurons in the second hidden layer of the neural network is [1, 100].

[0088] An SCR system concentration prediction model is built using an LSTM neural network, implemented using existing technology, and the other layers of the LSTM neural network are not limited.

[0089] In practical applications, step three involves training the SCR system concentration prediction model corresponding to each hyperparameter combination in the population matrix X using the constructed sample set. The optimal hyperparameter combination is then searched for within the population matrix X and implemented using the SSA algorithm. Other optimization algorithms, not limited to the SSA algorithm, can also be used.

[0090] Verification experiment:

[0091] To verify the effectiveness of the method for predicting NOx concentration in diesel engine SCR systems described in this invention, the number of hyperparameters to be optimized was determined to be 4, i.e., the search dimension was 4. The four hyperparameters were the number of model training iterations, the learning rate, and the number of neurons in the first and second hidden layers of the LSTM neural network. The threshold range for the number of model training iterations was [10, 100]; the threshold range for the learning rate was [0.001, 0.01]; the threshold range for the number of neurons in the first hidden layer of the neural network was [1, 100]; and the threshold range for the number of neurons in the second hidden layer of the neural network was [1, 100]. The following figures show the changes in relevant data during the process of determining the optimal combination of hyperparameters for the SCR system concentration prediction model:

[0092] Figure 2 The graph shows the change in fitness values ​​corresponding to the initial optimal hyperparameter combination during the iteration process. It can be seen from the graph that as the number of iterations increases during the search process, the fitness value continuously decreases, indicating that the SCR system concentration prediction model corresponding to this hyperparameter combination is better. At the 17th iteration, the fitness value no longer changes and reaches its minimum. The initial optimal hyperparameter combination corresponding to this iteration count is considered the optimal hyperparameter combination. Figure 2 It can be seen that for the 56515 engine, the fitness value tends to stabilize from the 17th iteration onwards, with a stable value of 0.000365. This indicates that the optimized SCR system concentration prediction model converges extremely easily.

[0093] Figure 3 The curve showing the change in the learning rate during the optimization process; Figure 4 Optimize the curve for the number of training iterations of the model; Figure 5 Optimize the curve for the number of neurons in the first hidden layer; Figure 6 The curve for optimizing the number of neurons in the second hidden layer.

[0094] The optimal hyperparameter combination for the best SCR system concentration prediction model is shown in Table 1:

[0095] Table 1 Optimal Hyperparameter Combinations

[0096]

[0097] As shown in Table 1, the four hyperparameters of the SCR system concentration prediction model are: learning rate of 0.00499, training times of 89, number of neurons in the first hidden layer of 20, and number of neurons in the second hidden layer of 61.

[0098] To more clearly show the fitting trend of the exhaust, the first 200 data points of the engine were selected for display. Figure 7 This is a comparison chart between the predicted values ​​of NOx concentration in downstream SCR system emissions using the SCR system concentration prediction model of this invention and the actual collected values ​​of NOx concentration in downstream SCR system emissions.

[0099] Figure 8 This is a comparison chart showing the fitted distribution of predicted and actual NOx emissions using the SCR system concentration prediction model of this invention. Figure 8 The straight line in the diagram serves as a baseline for evaluating whether the predicted value and the actual value are equal. When the predicted value and the actual value are equal, they fall on the straight line. Figure 8 The fact that most of the predicted values ​​are close to the baseline indicates that the error between the measured and predicted values ​​of this invention is very small, proving the effectiveness of the prediction of this invention.

[0100] Table 2 Evaluation results of the present invention model and existing models.

[0101]

[0102] As shown in Table 2, when predicting NOx concentration in the SCR system exhaust using data from the 56515 engine, the root mean square errors (RMSE) of the SCR system concentration prediction model, LSTM, and MLP of this invention were 10.577, 10.702, and 12.581, respectively. The SCR system concentration prediction model's RMSE differed from the MLP by 2.004, a reduction of 15.9%. Compared to LSTM, its RMSE differed by 0.125, a reduction of 1.17%. The mean absolute errors (MAE) of the SCR system concentration prediction model, LSTM, and MLP of this invention were 2.199, 2.248, and 6.548, respectively. The SCR system concentration prediction model's MAE differed from the MLP by 4.349, a reduction of 66.4%; compared to LSTM, its MAE differed by 0.049, a reduction of 2.48%. The coefficient of determination (R²) of the SCR system concentration prediction model, LSTM, and MLP prediction models of this invention... 2 The coefficients of determination R0 for the SSA-LSTM model are 0.991, 0.989, and 0.985, respectively. Compared with the MLP and LSTM models, the R0 of the SSA-LSTM model is significantly higher. 2 It increased by 3.35% and 0.2%; among which, the coefficient of determination R 2 Used to evaluate the accuracy of the model.

[0103] Using data from two different engines, the 56515 engine and the 56515 engine, the NOx concentration prediction model for the SCR system was validated. Results from the 56515 engine were compared with those from the 56515 engine. The main objective was to validate the performance of the SCR system concentration prediction model on different datasets. Figure 9The graph shows a comparison between predicted and actual NOx values ​​for different engines.

[0104] The predictive performance evaluation indicators of the engine are shown in Table 3:

[0105] Table 3 Evaluation results of each model

[0106]

[0107] Using data from engines 56718 and 66439 for SCR system NOx concentration prediction, after optimizing the SCR system concentration prediction model using the method of this invention, the RMSE and MAE of engines 56718 and 66439 increased compared to engine 56515, indicating a decrease in accuracy. For Figure 9 It is readily apparent that the data points from the two different engine models used in the test model are not concentrated on either side of or on the baseline; some points are far from the line, thus reducing the prediction accuracy. However, the overall accuracy remains around 95%, indicating good prediction performance. This demonstrates that the optimized SCR system concentration prediction model of this invention has a certain generalization ability. It provides a new approach and method for predicting NOx emissions. Furthermore, the SCR system concentration prediction model can fit the actual sensor measurements well on different engines. Deploying the established model in the cloud allows it to replace sensors in achieving adaptive urea injection adjustment and emission monitoring functions; simultaneously, simplifying the aftertreatment system sensors improves fuel economy.

[0108] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for predicting NOx concentration in the exhaust of a diesel engine SCR system, characterized in that, The method includes the following steps: Step 1: Construct a dataset for NOx prediction based on SCR system data collected under real historical operating conditions; and add a concentration output label to each sample in the dataset. Each sample consists of 8 parameters: diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, average SCR ammonia storage ratio, average urea pressure, reducing agent injection quantity, and exhaust flow rate. The concentration output label for each sample is the NOx concentration output value of the SCR system at the corresponding time point; Step 2: Build an SCR system concentration prediction model using an LSTM neural network; Step 3: Determine the hyperparameters to be optimized in the SCR system concentration prediction model, and construct a population matrix in the search space using the hyperparameters to be optimized. Among them, the population matrix Including Groups of hyperparameters, where the types of hyperparameters in each group are identical; each group of hyperparameters includes There are 10 hyperparameters, and each hyperparameter corresponds to a type. Population Matrix ; Indicates the first The location of a set of hyperparameter combinations in the search space. ; Indicates the first Group of hyperparameter combinations in the first The position of the dimension ; Then use the constructed sample set to analyze the population matrix. The SCR system concentration prediction model was trained under each hyperparameter combination, starting from the population matrix. The optimal combination of hyperparameters is searched, and the SCR system concentration prediction model under the optimal combination of hyperparameters is taken as the optimal SCR system concentration prediction model. From the population matrix The specific process of searching for the optimal combination of hyperparameters includes: Step 31: Set the iteration threshold The population matrix is ​​initialized within the search space formed by the thresholds of each hyperparameter to be optimized. ; Step 3.2 The initial value is 1, the first In the next iteration, the constructed sample set is used to refine the population matrix. The SCR system concentration prediction model corresponding to each hyperparameter combination was trained to obtain the population matrix. fitness function vector in, For the first The fitness function corresponding to a combination of hyperparameters; ; for The value; Step 33, in the At the nth iteration, the fitness function vector The combination of hyperparameters corresponding to the minimum fitness function value is taken as the initial optimal hyperparameter combination; Steps 3 and 4: Judgment Has the iteration threshold been reached? If the result is yes, proceed to step three and six; if the result is no, proceed to step three and five. Step 35: Update the population matrix within the search space. ,make Proceed to step three two; Step 36: Select the minimum value of the fitness function value among all the initial optimal hyperparameter combinations obtained under all iterations, and take the initial optimal hyperparameter combination corresponding to the minimum fitness function value as the optimal hyperparameter combination. Step 4: Use the current diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, SCR ammonia average storage ratio, urea pressure average value, reductant injection demand, and exhaust flow rate as inputs to the optimal SCR system concentration prediction model. Perform concentration prediction through the optimal SCR system concentration prediction model and output the current SCR system exhaust NOx concentration prediction value.

2. The method for predicting NOx concentration in diesel engine SCR systems according to claim 1, characterized in that, Step one, the specific process of constructing the dataset used for NOx prediction includes: Step 1: Collect various types of parameters of the SCR system under real historical operating conditions within a preset time period; Step 1 and Step 2: Filter the various types of parameters of the SCR system collected at each time point, and select the 8 parameters corresponding to each time point, namely: diesel engine NOx concentration output value, coolant temperature output value, SCR inlet temperature output value, SCR outlet temperature output value, SCR ammonia average storage ratio, urea pressure average value, reducing agent injection quantity and exhaust flow rate. Step 13: Preprocess the 8 parameters corresponding to each time point to remove null values ​​and outliers, and normalize the 8 parameters corresponding to each time point after preprocessing. Then, use the 8 parameters corresponding to each time point after normalization as a sample.

3. The method for predicting NOx concentration in diesel engine SCR systems according to claim 1, characterized in that, In step three, update the population matrix within the search space. At that time, population matrix middle The update methods include: in, For the current iteration number, the th The optimal position in dimension, and the fitness function value corresponding to this optimal position is the fitness function vector. The minimum value in; For the current iteration number, the th The worst position in dimension, and the fitness function value corresponding to this worst position is the fitness function vector. The maximum value in; These are random numbers generated according to a normal distribution with a mean of 0 and a variance of 1. The fitness function vector at the current iteration number. The maximum value of the fitness function in the data; The fitness function vector at the current iteration number. The minimum value of the fitness function in the model. For coefficients; It is a random number between [-1, 1].

4. The method for predicting NOx concentration in diesel engine SCR systems according to claim 1, characterized in that, In step three, the search space is composed of the threshold values ​​of each hyperparameter.

5. The method for predicting NOx concentration in diesel engine SCR systems according to claim 1, characterized in that, When the value of is 4, the four hyperparameters in each hyperparameter combination are the number of training iterations, the learning rate, and the number of neurons in the first and second hidden layers of the LSTM neural network.

6. The method for predicting NOx concentration in diesel engine SCR systems according to claim 5, characterized in that, The threshold range for the number of training iterations is [10, 100]; the threshold range for the learning rate is [0.001, 0.01]; the threshold range for the number of neurons in the first hidden layer of the neural network is [1, 100]; and the threshold range for the number of neurons in the second hidden layer of the neural network is [1, 100].

7. The method for predicting NOx concentration in diesel engine SCR system exhaust according to claim 1, characterized in that, In step three, the number of hyperparameters to be optimized is the same as the dimension of the search space.

Images