A method and system for detecting nitrogen oxide by using a Latin hypercube sampling dynamic optimization LSTM for urban solid waste incineration

Abstract

The application discloses a kind of Latin hypercube sampling dynamic optimization LSTM nitrogen oxide detection method and system for municipal solid waste incineration: according to the mechanism of solid waste incineration process, process variables closely related to NOx are selected as the original feature set, and the most relevant N variables are selected as the input variables of the NOx concentration prediction model;After data preprocessing, the time series data is obtained by using the LSTM network;A dynamic optimization LSTM network based on Latin hypercube sampling initialization search space is constructed, and the hyperparameters of the LSTM network are automatically optimized and dynamically adjusted by combining the Bayesian optimization framework to obtain the NOx concentration prediction model;The output result is de-normalized to obtain the predicted value of nitrogen oxide NOx.The application significantly improves the global search capability and convergence speed of Bayesian optimization, and improves the NOx prediction accuracy and robustness of the LSTM model in the multi-condition solid waste incineration scene.

Description

Technical Field

[0001] This invention belongs to the field of urban solid waste treatment and intelligent pollutant detection technology, and in particular relates to a dynamic optimization LSTM method and system for detecting nitrogen oxides using Latin hypercube sampling for urban solid waste incineration. Background Technology

[0002] With the accelerating pace of urbanization, Chinese cities are facing a growing problem of solid waste, with the amount of urban solid waste generated increasing year by year. During solid waste incineration, harmful pollutants such as nitrogen oxides (NOx) are produced. NOx is one of the key pollutants emitted from urban solid waste incineration, and its emissions not only cause air pollution but also pose a potential threat to human health and the ecological environment. Therefore, real-time, accurate detection and prediction of NOx concentrations during urban solid waste incineration are crucial for timely and effective control and management measures.

[0003] Long Short-Term Memory (LSTM) networks possess powerful time-series data processing capabilities, enabling them to capture the dynamic relationships between various variables and NOx during solid waste incineration, making them suitable for NOx concentration prediction. However, the hyperparameters of standard LSTM networks (such as the number of hidden units, learning rate, and number of training rounds) typically rely on manual experience for setting, resulting in strong subjectivity, low optimization efficiency, and susceptibility to overfitting or underfitting. This makes them ill-suited to handling the multi-condition fluctuations and strong nonlinear characteristics of solid waste incineration. Existing methods often employ random or grid sampling in the initial sampling phase, which can easily lead to uneven distribution of sampling points, reduced optimization efficiency, and a tendency to get trapped in local optima. Summary of the Invention

[0004] The purpose of this invention is to propose a dynamic optimization strategy that combines Latin hypercube sampling and Bayesian optimization. By generating initial hyperparameter points that uniformly cover the search space through Latin hypercube sampling, the global search capability and convergence speed of Bayesian optimization are significantly improved, thereby further enhancing the NOx prediction accuracy and robustness of the LSTM model in multi-condition solid waste incineration scenarios.

[0005] This invention provides the following technical solution: A dynamic optimization LSTM method for nitrogen oxide detection using Latin hypercube sampling for urban solid waste incineration includes the following steps: Based on the mechanism of solid waste incineration, process variables closely related to NOx were selected as the original feature set, and the N most relevant variables were selected as input variables for the NOx concentration prediction model. The selected N variables are preprocessed to form time series data; Based on the preprocessed time series data, a dynamically optimized LSTM network based on Latin hypercube sampling initialization search space is constructed. Then, the hyperparameters of the LSTM network are automatically optimized and dynamically adjusted using a Bayesian optimization framework to obtain a NOx concentration prediction model. The output of the NOx concentration prediction model was inversely normalized to obtain the predicted NOx value.

[0006] Preferred methods for selecting the most relevant N variables include mutual information, maximum correlation, and minimum redundancy.

[0007] Preferred methods for preprocessing the selected N variables include: Savitzky-Golay filtering was used to smooth the filtered data; based on The criteria exclude data with NOx concentrations exceeding three standard deviations. Min-max normalization is used to unify data units.

[0008] Preferably, the LSTM network mainly consists of an input gate, a forget gate, an output gate, and cell states; First, based on the output value from the previous time step... and the input value at this moment Enter the Gate of Oblivion and decide which information to keep or discard; Then, the output value from the previous time step. and the input value at this moment Enter the input field again to obtain the information that determines which update to make. and the current cell state ; Finally, the final cell state is obtained from the output values ​​of the forget gate and the input gate. and output information And pass it down as the input value for the next time step.

[0009] Preferably, the method for constructing a dynamically optimized LSTM network based on Latin hypercube sampling initialization of the search space, and combining it with a Bayesian optimization framework to automatically optimize and dynamically adjust the hyperparameters of the LSTM network includes: Construct training and validation datasets; Set the hyperparameters and optimization range for Bayesian optimization; The initial hyperparameter samples are generated using the Latin hypercube sampling method; For each sampling point, the parameters of the sample are substituted into the LSTM network and the mean square error of the validation set under the current hyperparameter configuration is calculated as the objective function. The RBF kernel function is used to construct a probabilistic model of the objective function: , In the formula, This is the bandwidth parameter of the Gaussian kernel function, with a value of 1.0. The Euclidean distance between two points is: The next most worthwhile hyperparameter point to evaluate is selected by improving the expectation-maximization function; Repeat the above steps to select the hyperparameter combination that minimizes the objective function value as the optimal hyperparameter; The optimal hyperparameters are assigned to the trained Latin hypercube sampling dynamic optimization LSTM network to construct the NOx concentration prediction model.

[0010] This invention also provides a Latin hypercube sampling dynamic optimization LSTM nitrogen oxide detection system for urban solid waste incineration, which applies the above detection method and includes a variable unit, a data preprocessing unit, a prediction model construction unit, and a concentration prediction unit. The variable unit is used to select process variables closely related to NOx as the original feature set based on the solid waste incineration process mechanism, and to screen out the N most relevant variables as input variables for the NOx concentration prediction model. The data preprocessing unit is used to preprocess the selected N variables to form time series data; The prediction model building unit is used to construct a dynamically optimized LSTM network based on the preprocessed time series data and the initial search space of the Latin hypercube sampling. Then, the hyperparameters of the LSTM network are automatically optimized and dynamically adjusted by combining the Bayesian optimization framework to obtain the NOx concentration prediction model. The concentration prediction unit is used to perform inverse normalization on the output of the NOx concentration prediction model to obtain the predicted value of nitrogen oxides (NOx).

[0011] The beneficial effects of this invention are as follows: This invention provides a dynamic optimization LSTM method and system for nitrogen oxide detection using Latin hypercube sampling for urban solid waste incineration, which has the following advantages compared to traditional methods: (1) In view of the problems that the mechanism of urban solid waste incineration is complex, the operating conditions are varied, many process variables are coupled with each other, and the nonlinear relationship between NOx and other variables is difficult to accurately describe, this invention constructs an LSTM neural network model. By utilizing its powerful long-term dependency capture capability, it realizes high-precision prediction of NOx concentration at future time. The model has strong adaptability and can adapt to the dynamic characteristics of solid waste incineration process.

[0012] (2) To address the problems of LSTM neural network hyperparameters relying on manual experience for adjustment, which are cumbersome and have uncertain results, this invention introduces a Bayesian optimization algorithm to achieve automatic hyperparameter optimization. Unlike conventional Bayesian optimization methods, this invention further employs Latin hypercube sampling to uniformly cover and initialize the initial search space, effectively avoiding sample clustering and search blind spots caused by random sampling, improving the global search capability and convergence speed of Bayesian optimization, and forming a unique Latin hypercube sampling dynamic optimization strategy.

[0013] (3) Through the dynamic combination of Latin hypercube sampling and Bayesian optimization, the present invention can automatically obtain the optimal hyperparameter combination (number of hidden units, learning rate, number of training rounds) of the LSTM network without manual intervention, which significantly improves the NOx prediction accuracy and robustness of the model under complex nonlinear and multi-condition conditions, while avoiding the tediousness and uncertainty of traditional manual parameter tuning. Compared with the conventional BO-LSTM method, it has better optimization efficiency and prediction performance. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly described 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. Figure 1 This is a schematic diagram of the nitrogen oxide detection method of the present invention; Figure 2 This is the LSTM neural network topology diagram of the present invention; Figure 3 This is the RMSE curve of the NOx prediction method of the present invention during the training phase; Figure 4 This is a graph showing the results of the NOx prediction method of this invention on the test set; Figure 5 This is an error curve of the NOx prediction method of the present invention on the test set; Figure 6 This is a graph showing the change in hyperparameter magnitudes during the training phase of the NOx prediction method of this invention. Detailed Implementation

[0015] 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.

[0016] Example 1 This invention provides a dynamic optimization LSTM method for NOx detection based on Latin hypercube sampling for urban solid waste incineration. The overall technical approach is as follows: First, based on the mechanism of solid waste incineration, key variables are screened using mutual information methods and the maximum correlation minimum redundancy algorithm, providing a foundation for subsequent detection accuracy. Then, the screened variables undergo data preprocessing, including Savitzky-Golay filtering, outlier removal, and normalization, to reduce the impact of noise and outliers on the model. Next, an LSTM network is constructed to model the time-series data. Finally, Latin hypercube sampling is introduced to initialize the Bayesian optimization search space, forming a dynamic optimization strategy to automatically optimize and dynamically adjust the LSTM network hyperparameters (number of hidden units, learning rate, number of training epochs). The network model is trained and validated using actual data from the solid waste incineration process, ultimately achieving efficient and accurate prediction of NOx concentration. The key to this technology lies in initializing the Bayesian-optimized search space through Latin hypercube sampling, enabling dynamic and automatic adjustment of the hyperparameters of the Long Short-Term Memory (LSTM) network. This improves the accuracy and robustness of nitrogen oxide prediction under complex nonlinear and multi-condition working conditions. The overall solution is as follows: Figure 1 As shown.

[0017] The following is in conjunction with the appendix Figure 1-6 To elaborate in detail: S1. Based on the mechanism of solid waste incineration, process variables closely related to NOx are selected as the original feature set, and the N most relevant variables are selected as input variables for the NOx concentration prediction model.

[0018] In this embodiment, the screening methods mainly include the mutual information (MI) method, the maximum correlation method, and the minimum redundancy method (mRMR): First, the similarity between any two variables in the dataset is calculated using the mutual information (MI) method: (1) In the formula, and Representing variables and The probability density function, Representing variables and The joint probability density function, , They represent the variables respectively and The integral variable.

[0019] Find the set of features in the original feature set that has the highest correlation with the final output NOx and the lowest correlation between the features.

[0020] Therefore, we first use the maximum correlation standard to select features that are highly correlated with NOx: (2) In the formula, Represents the set of all variables. Represents the first in the set of variables One variable, Represents the NOx variable. A correlation index indicating the relationship between a feature variable and a target variable. Indicates input variables Mutual information between NOx concentration and NOx concentration.

[0021] Then, the minimum redundancy standard form is used to remove redundant vectors from the variable set: (3) In the formula, Represents the set of all variables. Represents the first in the set of variables One variable, Indicates the degree of redundancy between features. Indicates input variables and Mutual information between them.

[0022] Combining equations (2) and (3), the formula for calculating mRMR is: (4) In the formula This represents the mRMR feature evaluation function. A correlation index representing the relationship between input features and the target variable. An indicator representing the redundancy between input features and other features.

[0023] Finally, the results were sorted in descending order, and the top 10 variables were selected as input variables for the NOx concentration prediction model. These included: the temperature of the right outer side of grate 1-1 in the combustion section, the flue gas temperature on the right side of the primary combustion chamber, the average temperature of the primary combustion chamber, the temperature of the left inner side of grate 1-2 in the combustion section, the flue gas temperature on the left side of the primary combustion chamber, the temperature on the right side of the primary combustion chamber, the flue gas temperature on the right side of the primary combustion chamber, the temperature of the left inner side of grate 1-1 in the combustion section, the air flow rate on the right side of the grate in the burnout section, and the air flow rate of the left section 1-2 of the grate in the combustion section.

[0024] S2. Perform data preprocessing on the selected N variables to form time series data.

[0025] In this embodiment, data preprocessing mainly includes data cleaning, noise reduction, outlier removal, and normalization. The specific methods are as follows: S2.1. Smooth the filtered data using Savitzky-Golay filtering: (5) In the formula, This is half the width of the window, and its value is 5. These are the Sgolay filter coefficients, whose values ​​are determined by the window length and the third-order local polynomial least squares fitting, and are fixed weight coefficients. Represented as the first The first variable sequence of the nth variable sequence The original data at each sampling time point This is the filtered result for this sampling point, where The index represents the discrete sampling time and indicates the center position of the current sliding window. This is the offset within the window, and its value range is... .

[0026] S2.2. Remove outliers, based on The criteria exclude data with NOx concentrations exceeding three standard deviations: (6) In the formula, Indicates the first NOx sequence One value, and Let represent the mean and standard deviation of the NOx sequence, respectively; satisfying equation (6) These are considered outliers and removed.

[0027] S2.3. Since each variable has a different unit, data normalization is required. Min-max normalization is used to unify the data units: (7) In the formula, The first of the 10 selected variables One variable, , These are the minimum and maximum values ​​of the variable, respectively. This is the normalized data.

[0028] S3. Based on the preprocessed time series data, a dynamically optimized LSTM network based on Latin hypercube sampling initialization search space is constructed. Then, the hyperparameters of the LSTM network are automatically optimized and dynamically adjusted using a Bayesian optimization framework to obtain a NOx concentration prediction model.

[0029] S3.1. In this embodiment, the LSTM network mainly consists of an input gate, a forget gate, an output gate, and cell states: First, let's look at the output value from the previous time step. and the input value at this moment Enter the forget gate to decide which information to keep or discard; then, the output value from the previous time step... and the input value at this moment Enter the input field again to obtain the information that determines which update to make. and the current cell state Finally, the final cell state is obtained from the output values ​​of the forget gate and the input gate. and output information And pass it down as the input value for the next time step: Forgotten Gate: (8) Input Gate: (9) Cell state: (10) (11) Output gate: (12) (13) In the formula, Represents the sigmoid activation function , Represents the tanh activation function ; , , , These are the weight matrices for the forget gate, input gate, cell state, and output gate, respectively. , , , These are the bias vectors for the forget gate, input gate, cell state, and output gate, respectively. , , , These are the output matrices for the forget gate, input gate, cell state, and output gate, respectively. The output of the LSTM at time t; This represents element-wise multiplication within a matrix. (LSTM neural network topology diagram reference) Figure 2 As shown.

[0030] S3.2. Construct training and validation datasets based on the data, and then select the first... Data as training dataset ,back Data as a validation dataset .

[0031] S3.3. Set the hyperparameters and optimization range for Bayesian optimization. Select three hyperparameters: the number of LSTM hidden units, the learning rate, and the number of training epochs. The specific ranges are as follows: (14) in, The number of hidden layer units in the LSTM. To find the lower bound for the number of hidden units, To find the upper limit for the number of hidden units; For learning rate, This is the lower bound for finding the optimal learning rate. This represents the upper limit for optimizing the learning rate; For the number of training rounds, To find the lower bound of the number of training rounds, This is the upper limit for the number of training iterations.

[0032] S3.4. Initial hyperparameter samples were generated using the Latin Hypercube Sampling (LHS) method. A total of [number] samples were generated. There are 3 sample points, and each sample point is represented as follows: (15) Each sample point For inclusion , , The vector, This indicates transpose, specifically as follows: (16) In the formula, , , Representing the first The values ​​of the number of hidden units, learning rate, and number of training epochs for each sample point are used. The generated sample points also need to be mapped to the real hyperparameter space. (17) (18) (19) In the formula They represent the first Number of hidden layer units in each LHS sample point Learning rate Number of training rounds exist Normalized sampled values ​​for the interval.

[0033] S3.5. For each sampling point, substitute the above parameters into the LSTM network and calculate the mean squared error of the validation set under the current hyperparameter configuration as the objective function: (20) In the formula, This represents the array of predicted values ​​for the validation set. Represents the actual array of the validation set. This represents the number of valid samples in the validation set.

[0034] S3.6. The RBF kernel function is used to construct the probabilistic model of the objective function. The specific formula is as follows: (twenty one) In the formula, This is the bandwidth parameter of the Gaussian kernel function, with a value of 1.0. This represents the Euclidean distance between two points.

[0035] For a given training dataset The Gaussian process for each candidate point The predicted mean and variance of the Gaussian distribution are: (twenty two) (twenty three) In the formula, , They represent the first The predicted mean and variance of each candidate hyperparameter point. This represents the vector of observed values ​​of the objective function corresponding to the evaluated training data, where the i-th... The element is the first Mean square error calculated from the set of hyperparameters in the LSTM network It is the covariance matrix between the training data. It is the covariance between the candidate points and the training data. It is the covariance between candidate points. Let V be the noise variance, where Pick , where I is the identity matrix.

[0036] S3.7. Select the next most worthwhile hyperparameter point to evaluate using the Expectation-Maximization (EI) function: (twenty four) (25) In the formula, It is the currently known optimal objective value, that is, the point where the objective function value is minimized; This represents the set of objective function values ​​during the Bayesian optimization process; This represents the expected improvement value of the candidate hyperparameter point; and Let represent the predicted mean and variance of the candidate hyperparameter points, respectively; For standardized differences, This represents the cumulative distribution function of the normal distribution. It is the probability density function of a normal distribution; The next most worthwhile point to evaluate using EI is: (26) S3.8. Repeat S3.4-S3.7 a total of 10 times, and select the hyperparameter combination that minimizes the objective function value. (27) in, This represents the set of hyperparameter samples accumulated after repeating the Bayesian optimization process 10 times. It consists of the initial samples and the hyperparameter vectors added during the iteration process. Indicates the first among all hyperparameter samples The objective function value for each sample; This is the optimal hyperparameter configuration.

[0037] S3.9. Assign the optimal hyperparameters obtained in S3.7 to the trained Latin hypercube sampling dynamic optimization LSTM network to construct the NOx concentration prediction model.

[0038] In the formula, This represents the predicted NOx concentration at time t; It is based on a dynamically optimized LSTM network using Latin hypercube sampling. The input time-series feature matrix is, i.e., derived from... The matrix formed, where The step size of the LSTM model and Each in the sequence Both are 11-dimensional feature vectors, which can be represented as ,in The values ​​of the 10 relevant feature variables selected by mRMR at time i are given. Let be the NOx concentration at time i; This represents the optimal hyperparameter value.

[0039] S4. Input the test sample data into the NOx concentration prediction model, and perform inverse normalization on the output results to obtain the final predicted value of nitrogen oxides (NOx).

[0040] Example 2 This embodiment provides a verification of the effectiveness of the present invention, wherein the experimental data comes from the Gaoantun database.

[0041] First, we identified data segments that were relatively stable without drastic fluctuations, and finally selected 5,000 samples from these stable operating conditions.

[0042] Then, variables related to the mechanism of solid waste incineration were selected from all variables, specifically including: The 18 temperature-related variables are: flue gas temperature 1 on the left side of the primary combustion chamber, flue gas temperature 1 on the right side of the primary combustion chamber, flue gas temperature 2 on the left side of the primary combustion chamber, flue gas temperature 3 on the left side of the primary combustion chamber, flue gas temperature 3 on the right side of the primary combustion chamber, temperature on the right side of the primary combustion chamber, temperature on the left side of the primary combustion chamber, average temperature of the primary combustion chamber, temperature on the inner left side of grate 1-1 in the combustion section, temperature on the inner right side of grate 1-1 in the combustion section, temperature on the inner left side of grate 1-2 in the combustion section, temperature on the outer left side of grate 1-2 in the combustion section, temperature on the inner right side of grate 1-2 in the combustion section, temperature on the inner left side of grate 2-1 in the combustion section, temperature on the inner right side of grate 2-2 in the combustion section, and temperature on the inner right side of grate 2-2 in the combustion section.

[0043] The 13 variables related to air volume or air flow are: cumulative primary air volume of the furnace, air flow of the left 1 section of the drying grate, air flow of the left 1-1 section of the combustion section grate, air flow of the right 1-1 section of the combustion section grate, air flow of the left 1-2 section of the combustion section grate, air flow of the right 1-2 section of the combustion section grate, air flow of the left 2-1 section of the combustion section grate, air flow of the right 2-1 section of the combustion section grate, air flow of the left 2-2 section of the combustion section grate, air flow of the left grate of the burnout section grate, and cumulative secondary air volume of the furnace.

[0044] The two variables related to urea are: cumulative urea solvent supply flow rate and cumulative furnace urea solution volume.

[0045] The first 3000 samples were then used for training, and the last 2000 samples were used for testing. The LSTM neural network was trained on the training set and validated on the test set using the constructed Latin hypercube sampling dynamic optimization.

[0046] Specifically, the final hyperparameter optimization range and results after Latin hypercube sampling and Bayesian optimization are shown in Table 1: Table 1 In this embodiment, the RMSE curve of the NOx prediction method of the present invention during the training phase is as follows: Figure 3As shown, the X-axis represents the number of iterations, and the Y-axis represents the training RMSE value; the prediction results of the NOx prediction method on the test set are as follows. Figure 4 As shown, the X-axis represents the test sample, and the Y-axis represents the predicted NOx concentration. The dashed line represents the expected NOx prediction output, and the solid line represents the actual NOx prediction output. The NOx testing error is shown in the figure. Figure 5 As shown, the X-axis represents the number of test samples, and the Y-axis represents the prediction error; the graph shows the changes in the hyperparameters of the NOx prediction method during the training phase. Figure 6 As shown, the X-axis represents the number of Bayesian optimization rounds, and the Y-axis, from top to bottom, represents the number of hidden units, the learning rate, and the number of training rounds.

[0047] Example 3 This embodiment provides a Latin hypercube sampling dynamic optimization LSTM nitrogen oxide detection system for urban solid waste incineration, mainly including a variable unit, a data preprocessing unit, a prediction model construction unit, and a concentration prediction unit. Specifically: The variable unit is used to select process variables closely related to NOx as the original feature set based on the mechanism of solid waste incineration, and to screen out the N most relevant variables as input variables for the NOx concentration prediction model.

[0048] The data preprocessing unit is used to preprocess the selected N variables to form time series data.

[0049] The prediction model building unit is used to construct a dynamically optimized LSTM network based on the preprocessed time series data and the initial search space of the Latin hypercube sampling. The hyperparameters of the LSTM network are automatically optimized and dynamically adjusted using a Bayesian optimization framework to obtain the NOx concentration prediction model.

[0050] The concentration prediction unit is used to perform inverse normalization on the output of the NOx concentration prediction model to obtain the predicted value of nitrogen oxides (NOx).

[0051] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A dynamically optimized LSTM method for nitrogen oxide detection using Latin hypercube sampling for urban solid waste incineration, characterized in that, Includes the following steps: Based on the mechanism of solid waste incineration, process variables closely related to NOx were selected as the original feature set, and the N most relevant variables were selected as input variables for the NOx concentration prediction model. The selected N variables are preprocessed to form time series data; Based on the preprocessed time series data, a dynamically optimized LSTM network based on Latin hypercube sampling initialization search space is constructed. Then, the hyperparameters of the LSTM network are automatically optimized and dynamically adjusted using a Bayesian optimization framework to obtain a NOx concentration prediction model. The output of the NOx concentration prediction model was inversely normalized to obtain the predicted NOx value.

2. The method according to claim 1, characterized in that, Methods for selecting the N most relevant variables include mutual information, maximum correlation, and minimum redundancy.

3. The method according to claim 1, characterized in that, Methods for preprocessing the selected N variables include: Savitzky-Golay filtering was used to smooth the filtered data; based on The criteria exclude data with NOx concentrations exceeding three standard deviations. Min-max normalization is used to unify data units.

4. The method according to claim 1, characterized in that, An LSTM network mainly consists of an input gate, a forget gate, an output gate, and cell states; First, based on the output value from the previous time step... and the input value at this moment Enter the Gate of Oblivion and decide which information to keep or discard; Then, the output value from the previous time step. and the input value at this moment Enter the input field again to obtain the information that determines which update to make. and the current cell state ; Finally, the final cell state is obtained from the output values ​​of the forget gate and the input gate. and output information And pass it down as the input value for the next time step.

5. The method according to claim 4, characterized in that, Methods for constructing dynamically optimized LSTM networks based on Latin hypercube sampling initialization of the search space, and for automatically optimizing and dynamically adjusting the hyperparameters of the LSTM network using a Bayesian optimization framework, include: Construct training and validation datasets; Set the hyperparameters and optimization range for Bayesian optimization; The initial hyperparameter samples are generated using the Latin hypercube sampling method; For each sampling point, the parameters of the sample are substituted into the LSTM network and the mean square error of the validation set under the current hyperparameter configuration is calculated as the objective function. The RBF kernel function is used to construct a probabilistic model of the objective function: ， In the formula, This is the bandwidth parameter of the Gaussian kernel function, with a value of 1.

0. The Euclidean distance between two points is: The next most worthwhile hyperparameter point to evaluate is selected by improving the expectation-maximization function; Repeat the above steps to select the hyperparameter combination that minimizes the objective function value as the optimal hyperparameter; The optimal hyperparameters are assigned to the trained Latin hypercube sampling dynamic optimization LSTM network to construct the NOx concentration prediction model.

6. A dynamic optimization LSTM nitrogen oxide detection system using Latin hypercube sampling for urban solid waste incineration, characterized in that, It includes a variable unit, a data preprocessing unit, a prediction model building unit, and a concentration prediction unit; The variable unit is used to select process variables closely related to NOx as the original feature set based on the solid waste incineration process mechanism, and to screen out the N most relevant variables as input variables for the NOx concentration prediction model. The data preprocessing unit is used to preprocess the selected N variables to form time series data; The prediction model building unit is used to construct a dynamically optimized LSTM network based on the preprocessed time series data and the initial search space of the Latin hypercube sampling. Then, the hyperparameters of the LSTM network are automatically optimized and dynamically adjusted by combining the Bayesian optimization framework to obtain the NOx concentration prediction model. The concentration prediction unit is used to perform inverse normalization on the output of the NOx concentration prediction model to obtain the predicted value of nitrogen oxides (NOx).

Images