A method for predicting co emissions from an mswi process

Abstract

The application provides a MSWI process CO emission prediction method, which firstly carries out particle design oriented to reduction depth features and LSTM optimization, can adaptively determine a feature selection threshold range according to modeling data characteristics, secondly inputs deep features after nonlinear feature extraction by using super one-dimensional convolution into LSTM to construct a prediction model, updates convolution layers and LSTM hyperparameters based on a loss function, takes generalization performance of the model as a fitness function of an optimization algorithm, and finally adopts a particle swarm optimization (PSO) algorithm to adaptively select features and hyperparameters according to data characteristics. The MSWI process CO emission prediction method provided by the application can realize CO emission prediction in the MSWI process based on reduction depth features and LSTM optimization.

Description

Technical Field

[0001] This invention relates to the field of urban solid waste incineration technology, and in particular to a method for predicting CO emissions from the MSWI process. Background Technology

[0002] Carbon monoxide (CO), one of the toxic gases produced during MSWI processes, is difficult to detect by the senses. However, when it binds to hemoglobin, it hinders blood oxygen transport and can even cause myocardial infarction in severe cases, thus requiring strict control. Furthermore, it is directly related to dioxins (DXN), which cause the "not in my backyard" (NIMBY) effect in incineration plant construction. DXN is a "century-old poison" with strictly limited emissions internationally, exhibiting cumulative and highly toxic properties. Currently, achieving real-time, accurate, and low-cost detection of DXN emissions is a pressing problem for the industry. In contrast, CO can be detected in real-time using a CEMS (Continuous Emission Monitoring System). Due to the significant spike-pulse phenomenon in CO emission concentrations, the internationally accepted standard is based on the half-hour average. Research indicates that achieving intelligent optimization control and green production in complex industrial processes requires real-time monitoring of parameters such as product quality, energy and material consumption, and pollution emissions. For MSWI processes, predicting CO emissions can indirectly support the long-term, high-cost detection of DXN emissions, thereby assisting in the optimized control of MSWI processes.

[0003] From the perspective of generation mechanisms, CO emission concentrations are correlated with numerous process variables in the MSWI process, and these correlations vary under different operating conditions. This makes the high-dimensionality of the modeling data a primary challenge in CO prediction modeling. Processing high-dimensional features requires extensive computation; feature selection is an effective strategy for reducing dimensionality, removing irrelevant data, and improving learning accuracy. For complex industrial processes, only domain experts with extensive prior knowledge can select relevant process variables, but this approach also involves a degree of subjectivity and arbitrariness. Current research has not addressed how to determine the threshold for feature selection based on data characteristics, nor has it considered the deep-seated correlations between features. Therefore, designing a CO emission prediction method for the MSWI process is essential. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting CO emissions during the MSWI process, which can predict CO emissions during the MSWI process based on reduced depth features and LSTM optimization.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for predicting CO emissions from an MSWI process includes the following steps:

[0007] Step 1: Construct an optimization problem for CO emission prediction based on reduced deep features and LSTM optimization;

[0008] Step 2: For the optimization problem, conduct particle design oriented towards reducing deep features and LSTM optimization;

[0009] Step 3: Design a fitness function for optimization problems that is oriented towards reducing deep features and LSTM optimization;

[0010] Step 4: For the optimization problem, perform a PSO-based optimization process.

[0011] Optionally, in step 1, an optimization problem for CO emission prediction based on reduced deep features and LSTM optimization is constructed, specifically as follows:

[0012]

[0013]

[0014] In the formula, f(X) is the performance index of the model, which is used as the fitness function for optimization. 1DCNN-LSTM (·) represents a model based on 1DCNN and LSTM. Input the original sample dataset. For X ori The normalized dataset, X sel For MI feature selection f MI (·) is the feature dataset, and F is the test set in f 1DCNN-LSTM The generalization performance on (·) is used as the fitness f(X), X exa This is the feature dataset after feature extraction using 1DCNN, where n is the number of samples (N in total), and m is the number of convolutional layers (M in total). ori For X ori The dimension of M sel For X sel The dimension of M exa For X exa The dimension, θ numFilters and θ convlayers These represent the number of convolutional kernels and the number of convolutional layers, respectively.

[0015] Optionally, in step 2, for the optimization problem, particle design is performed for reducing depth features and LSTM optimization, specifically as follows:

[0016] Let θ be the decision variable for the particle. p ,Right now:

[0017] θ p ={θ MI ,θ numFilters ,θ convlayers,θ Epochs ,θ leamingrate ,θ dropout} (2)

[0018] In the formula, θ MI For the feature selection threshold based on MI, θ Epochs θ learningrate and θ dropout The MI value is calculated based on the modeling data, taking into account the number of iterations, learning rate, and dropout rate of the LSTM model. and As shown below:

[0019]

[0020]

[0021] In the formula, and These are the maximum and average values ​​of MI, respectively. for The i-th feature in y nom To normalize the CO concentration, each characteristic is related to y. nom The MI value between them is denoted as I. MI Based on the characteristics of the data, a suitable threshold range for modeling the data is obtained, i.e. Based on the set θ MI Obtain M sel For each feature, the hyperparameters of subsequent 1DCNN feature extraction should satisfy the following relationship: M sel ≥θ numFilters ≥θ convlayers The hyperparameters θ of the LSTM are determined based on empirical terms. Epochs θ learningrate and θ dropout The upper and lower limits.

[0022] Optionally, in step 3, a fitness function is designed for the optimization problem, focusing on reducing deep features and optimizing LSTM. This includes the following steps:

[0023] Step 301: For the original dataset {X} ori Normalize y to obtain a standard dataset

[0024] Step 302: Perform MI-based feature selection and calculate the relationship between individual features in the process variables and CO concentration y. nom The MI value between, and then based on θ MI get M in sel From the feature vectors, obtain the feature dataset X. sel ;

[0025] Step 303: Perform feature extraction using 1DCNN, through parameter θ numFilters and θ convlayers Convolution is performed to obtain the reduced feature dataset X exa And input it into the LSTM model;

[0026] Step 304: Using parameter θ Epochs θ learningrate and θ dropout Build a prediction model and update the network parameters according to the loss function during the modeling process.

[0027] Optionally, in step 301, the original dataset {X} is... ori Normalize y to obtain a standard dataset Specifically:

[0028] X ori After normalization in, The calculation is shown in the following formula:

[0029]

[0030] In the formula, For X ori The i-th feature, x i,min and x i,max Let be the minimum and maximum values ​​of the i-th feature, respectively. Normalize the data to [0, 1]. exist

[0031]

[0032] Similarly, normalize y to y nom .

[0033] Optionally, in step 302, feature selection based on MI is performed, and the relationship between a single feature in the process variable and the CO concentration y is calculated. nom The MI value between, and then based on θ MI get M in sel From the feature vectors, obtain the feature dataset X. sel Specifically:

[0034] Calculate the relationship between a single characteristic in the process variables and CO concentration y nom To determine the MI value between sets, first, set the empty set X. sel Calculate using the following formula The i-th feature value of the n-th sample With y nom The MI values ​​between them are:

[0035]

[0036] In the formula, Let i be the i-th feature value of the n-th sample. For the value of the nth sample, For joint probability distribution, For the marginal probability distribution, calculate I. MI for:

[0037]

[0038] Based on the set θ MI get M in sel The dataset after feature vector reduction is:

[0039]

[0040] Optionally, in step 303, feature extraction using 1D CNN is performed, through parameter θ. numFilters and θ convlayers Convolution is performed to obtain the reduced feature dataset X exa And input it into the LSTM model, specifically:

[0041] 1DCNN contains m = 1, ..., M convolutional layers, with X1 as its input. sel X1 exists sel =X sel The convolution process of the first layer is as follows:

[0042] First, determine the kernel size as M. sel ×θ numFilters ;

[0043] Next, with a stride of s, the samples are convolved with the convolution kernel. The convolution process is as follows:

[0044]

[0045] In the formula, W k,1 b k,1 These are the convolution kernel and bias of the first layer, respectively. The input X for the first convolution layer sel The nth sample, This is the output of the nth sample from the first convolutional layer. The output is the result after the first convolutional layer has processed all samples. for:

[0046]

[0047] Similarly, the output X of the Mth layer exa for:

[0048]

[0049] Optionally, in step 304, the parameter θ Epochs θ learningrate and θ dropout The prediction model is constructed, and the network parameters are updated according to the loss function during the modeling process, specifically as follows:

[0050] Reduce the feature dataset X exa And input it into the LSTM prediction model. Wherein, h represents the nth sample input. n-1 and C n-1 h represents the output and state information of the LSTM hidden layer for the (n-1)th sample, respectively, and is used as the input for the next sample. n and C n The output and state information of the LSTM for the nth sample are respectively. The specific process during the initial sample calculation is as follows:

[0051] Define the empty set C n For memory units, forget gate f n Output F n The calculation is as follows:

[0052]

[0053] F n =f n ⊙C n (14)

[0054] In the formula, U f and W f Corresponding to h respectively n-1 and The weight, b f For bias, σ(·) represents the sigmoid activation function, and ⊙ represents the Hadamard product, used to calculate candidate values ​​for the current input state. for:

[0055]

[0056] In the formula, U c and W c Corresponding to h respectively n-1 , The weight, b c For bias, tanh(·) denotes the tanh activation function;

[0057] Calculate the input gate i n Output I n for:

[0058]

[0059]

[0060] In the formula, U i and W i Corresponding to h respectively n-1 and The weight, b i As a bias, the obtained value is stored in memory cell C. n ,get:

[0061] C n =F n +I n (18)

[0062] The first sample output C n The value is I n C n =I n ,F n =0, calculate the hidden layer output h of the nth sample. n Calculate the output gate o n for:

[0063]

[0064] h n =o n ⊙tanh(C n (20)

[0065] In the formula, U o and W o Corresponding to h respectively n-1 and The weight, b o As a bias, calculate the output value of the nth sample. for:

[0066]

[0067] Among them, W out The weights correspond to the hidden layers. Similarly, when the remaining samples are input, the above process is repeated, and the output is... for:

[0068]

[0069] The loss function e′ is:

[0070]

[0071] In the formula, e′ represents The error between the actual data y and the feedback, y n It is the actual value of the nth sample. Based on f 1DCNN-LSTM The predicted value of the nth sample output by (·) is first updated by updating the convolutional layer parameters as follows:

[0072]

[0073]

[0074] in, The updated convolutional kernel and bias are respectively used. Next, the LSTM parameters are updated, and a hidden state C exists. n and h n ,definition and The calculation is as follows:

[0075]

[0076]

[0077] Calculate the weights and biases of the forget gate, input gate, and output gate. The learning formula for the output gate weights is as follows:

[0078]

[0079] The weights are updated as follows:

[0080]

[0081] In the formula, This refers to the updated output gate weights.

[0082] Optionally, in step 4, a PSO-based optimization process is performed for the optimization problem, specifically as follows:

[0083] The modeling process is optimized based on the PSO algorithm, which includes population initialization, updating particle velocity and selecting decision variables, calculating fitness, updating individual particle optimality, profile and sample pool, and obtaining global optimum and model hyperparameters after reaching the number of iterations.

[0084] Population initialization is as follows: First, the number of particles P is... num Number of iterations N iter Update threshold N refresh The parameters, such as the upper and lower limits of the decision variables, are set, and then the result is generated by P. num A population of particles, with θ randomly initialized for each particle. p v p And calculate the particle's f(X), then initialize the particle's d. p Finally, with A, the particle's... With E P ;

[0085] Update the particle velocity and select decision variables, with the objective function being to maximize f(X), whose velocity v p The update formula is:

[0086]

[0087] In the formula, w inertia The inertial weights that affect the particle search step size, where c is the learning factor. They follow a uniform distribution between [0,1]. For the nth dimension of particle p, the particle position θ p The update formula is:

[0088]

[0089] Fitness is calculated by calculating R0 on the test set. 2 As a particle fitness metric, the specific process is as follows:

[0090]

[0091] The particle update formula is as follows: (Update individual particle optimality, file, and sample pool.)

[0092]

[0093] In the formula, In the individual optimal formula for a particle, `pop` represents the particle's information, and `popmin(site)` and `popmax(site)` represent the minimum and maximum thresholds for the particle's range at position `site`. If the individual particle is in the optimal iteration N... refresh If no update is performed after this step, the learning sample pool will be updated. The strategy is to set the update probability of each dimension of the particle's learning sample as follows: During the update, first, two particles are randomly selected from the population. Then, the individual optimality of the two particles is compared. After the competition, the better individual optimality is selected as the new learning example, as follows:

[0094]

[0095] Learning probability of particle p The update is as follows:

[0096]

[0097] As particle sorting rank p As the learning probability increases, the probability of updating the learned sample gradually increases within the range of 5% to 50%.

[0098] Calculating the global optimum: After initializing the population, the iterative optimization phase begins. First, the particle velocity v is updated. p Update the position θ of the particle's parameter decision variable. p After obtaining the hyperparameter values, feature reduction, model construction, and calculation of the evaluation index R are performed. 2 Using f(X) as the particle fitness value, update the individual optimality of the particle based on the fitness value, store the non-dominated solutions found in the population search into the archive, update the archive A, and calculate the individual optimal ranking rank of the particles. p And update its learning probability P c Therefore, for iteration N refresh After that, the particles whose individual optimal values ​​have not been updated are updated with learning examples. The particle with the largest f(X) value in the file is taken as the global optimal, and the optimal hyperparameters are obtained after decoding.

[0099] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The MSWI process CO emission prediction method provided by the present invention firstly employs particle design oriented towards reduced deep features and LSTM optimization, which can adaptively determine the feature selection threshold range based on the characteristics of the modeling data; secondly, the deep features extracted using nonlinear features through super-one-dimensional convolution are input into an LSTM to construct a prediction model, and the hyperparameters of the convolutional layer and LSTM are updated based on a loss function, using the generalization performance of the model as the fitness function of the optimization algorithm; finally, a particle swarm optimization (PSO) algorithm is used to adaptively select features and hyperparameters based on data characteristics. The rationality and effectiveness of the method are verified on the MSWI process CO dataset, enabling CO emission prediction in the MSWI process based on reduced deep features and LSTM optimization. Attached Figure Description

[0100] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0101] Figure 1 A schematic diagram of the urban solid waste incineration process;

[0102] Figure 2 This is a schematic diagram of the strategy flow of the MSWI process CO emission prediction method according to an embodiment of the present invention;

[0103] Figure 3a This is a diagram illustrating the changes in MI values;

[0104] Figure 3b This is a diagram illustrating the changes in the numFilters value;

[0105] Figure 3c This is a diagram illustrating the changes in convlayers values;

[0106] Figure 3d This is a diagram illustrating the changes in Epochs values;

[0107] Figure 3e This is a diagram illustrating the changes in the learning rate value;

[0108] Figure 3f This is a diagram illustrating the changes in dropout values;

[0109] Figure 4 For the CO dataset, the number of iterations and R 2 Relationship diagram;

[0110] Figure 5a A graph showing the predicted CO concentration for the training set;

[0111] Figure 5b A graph showing the predicted CO concentration for the validation set;

[0112] Figure 5c This is a predicted CO concentration curve for the test set. Detailed Implementation

[0113] The purpose of this invention is to provide a method for predicting CO emissions during the MSWI process, which can predict CO emissions during the MSWI process based on reduced depth features and LSTM optimization.

[0114] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0115] Urban solid waste incineration process such as Figure 1 As shown, research indicates that CO in G3 flue gas is directly related to DXN emission concentration. Hasberg et al. provided the following mapping relationship between DXN and CO:

[0116]

[0117] Where, ρ DXN ρ represents the concentration of DXN. CO k represents the CO concentration. CO Since CO is a constant, an increase in CO concentration indicates a greater degree of incomplete combustion, which may lead to an increase in DXN emission concentration. From an engineering practice perspective, CO emission concentration should be strictly controlled in industrial sites. Therefore, predicting CO emission concentration is essential for reducing pollution emissions from the MSWI process through optimized control.

[0118] like Figure 2 As shown in the embodiment of the present invention, the method for predicting CO emissions from the MSWI process includes the following steps:

[0119] Step 1: Construct an optimization problem for CO emission prediction based on reduced deep features and LSTM optimization;

[0120] Step 2: For the optimization problem, conduct particle design oriented towards reducing deep features and LSTM optimization;

[0121] Step 3: Design a fitness function for optimization problems that is oriented towards reducing deep features and LSTM optimization;

[0122] Step 4: For the optimization problem, perform a PSO-based optimization process.

[0123] In step 1, an optimization problem for CO emission prediction based on reduced deep features and LSTM optimization is constructed, specifically as follows:

[0124] The optimization problem for CO emission prediction based on reduced depth features and LSTM optimization is as follows:

[0125]

[0126]

[0127] In the formula, f(X) is the performance index of the model, which is used as the fitness function for optimization. 1DCNN-LSTM (·) represents a model based on 1DCNN and LSTM. Input the original sample dataset. For X ori The normalized dataset, X sel For MI feature selection f MI (·) is the feature dataset, and F is the test set in f 1DCNN-LSTM The generalization performance on (·) is used as the fitness f(X), X exa This is the feature dataset after feature extraction using 1DCNN, where n is the number of samples (N in total), and m is the number of convolutional layers (M in total). ori For X ori The dimension of M sel For X sel The dimension of M exa For X exa The dimension, θ numFilters and θ convlayers These represent the number of convolutional kernels and the number of convolutional layers, respectively.

[0128] The specific strategy diagrams for steps 2, 3, and 4 are as follows: Figure 2 As shown, Figure 2 In, θ pθ represents the particle decision variable and is also the particle position in PSO; MI This represents the feature selection threshold based on MI. and These represent the maximum and average values ​​of MI, respectively. This represents the prediction result input into the LSTM model after feature extraction via 1DCNN; θ Epochs θ learningrate and θ dropout This indicates the number of iterations, learning rate, and dropout rate of the LSTM model; f MI Enter f after (·) 1DCNN-LSTM The predicted value of (·); v p d represents the particle velocity in PSO; p A represents the optimal individual particle, and A represents the file; N iter The number of iterations for PSO optimization is represented by... Figure 2 The proposed strategy involves the following steps: First, selecting the MI threshold range based on data characteristics to obtain the features selected by MI; next, optimizing the feature extraction of 1DCNN; and further, optimizing the LSTM hyperparameters to achieve model construction. The specific steps are shown below.

[0129] In step 2, for the optimization problem, particle design is carried out for reducing deep features and LSTM optimization, specifically as follows:

[0130] Calculate based on modeling data and and take it as θ MI The upper and lower limits, based on the obtained θ MI The number of features M after feature selection is obtained. sel , such that θ numFilters With M sel Satisfy M sel ≥θ numFilters , and then θ numFilters ≥θ convlayers Further determine the hyperparameter θ of the LSTM Epochs θ learningrate and θ dropout The scope and detailed process are as follows:

[0131] Let θ be the decision variable for the particle. p ,Right now:

[0132] θ p ={θ MI ,θ numFilters ,θ convlayers ,θ Epochs ,θ leamingrate ,θ dropout} (3)

[0133] In the formula, there are a total of 6 variables, corresponding to the hyperparameters of the modeling process, θ MI For the feature selection threshold based on MI, θ Epochs θ learningrate and θ dropout The MI value is calculated based on the modeling data, taking into account the number of iterations, learning rate, and dropout rate of the LSTM model. and As shown below:

[0134]

[0135]

[0136] In the formula, and These are the maximum and average values ​​of MI, respectively. for The i-th feature in y nom To normalize the CO concentration, each characteristic is related to y. nom The MI value between them is denoted as I. MI Based on the characteristics of the data, a suitable threshold range for modeling the data is obtained, i.e. Based on the set θ MI Obtain M sel For each feature, the hyperparameters of subsequent 1DCNN feature extraction should satisfy the following relationship: M sel ≥θ numFilters ≥θ convlayers The hyperparameters θ of the LSTM are determined based on empirical terms. Epochs θ learningrate and θ dropout The upper and lower limits;

[0137] Each particle contains θ p v p f(X), learning example E P Individual optimal ranking p and learning probability Attributes, where: particle rank p Confirmed Impact E P Update; E P Guided particle v p The search direction and step size determine the particle θ p Update; based on θ p Calculate f(X), and update the individual optimality and ranking of particles using the value of f(X).

[0138] In step 3, for the optimization problem, a fitness function is designed for reducing deep features and optimizing LSTM, specifically including the following steps:

[0139] The change in the fitness function value is based on the particle's position θ. p The process of calculating the optimization objective defined by formula (2) in this invention will evaluate the performance index R of the test set. 2 Fitness as a particle;

[0140] Step 301: For the original dataset {X} ori Normalize y to obtain a standard dataset

[0141] Step 302: Perform MI-based feature selection and calculate the relationship between individual features in the process variables and CO concentration y. nom The MI value between, and then based on θ MI get M in sel From the feature vectors, obtain the feature dataset X. sel ;

[0142] Step 303: Perform feature extraction using 1DCNN, through parameter θ numFilters and θ convlayers Convolution is performed to obtain the reduced feature dataset X exa And input it into the LSTM model;

[0143] Step 304: Using parameter θ Epochs θ learningrate and θ dropout Build a prediction model and update the network parameters according to the loss function during the modeling process.

[0144] In step 301, the original dataset {X} is processed. ori Normalize y to obtain a standard dataset Specifically: X ori After normalization in, The calculation is shown in the following formula:

[0145]

[0146] In the formula, For X ori The i-th feature, x i,min and x i,max Let be the minimum and maximum values ​​of the i-th feature, respectively. Normalize the data to [0, 1]. exist

[0147] Similarly, normalize y to ynom ;

[0148] In step 302, feature selection based on MI is performed, and the relationship between individual features in the process variables and CO concentration y is calculated. nom The MI value between, and then based on θ MI get M in sel From the feature vectors, obtain the feature dataset X. sel Specifically:

[0149] Calculate the relationship between a single characteristic in the process variables and CO concentration y nom To determine the MI value between sets, first, set the empty set X. sel Calculate using the following formula The i-th feature value of the n-th sample With y nom The MI values ​​between them are:

[0150]

[0151] In the formula, Let i be the i-th feature value of the n-th sample. For the value of the nth sample, For joint probability distribution, For the marginal probability distribution, calculate I. MI for:

[0152]

[0153] Based on the set θ MI get M in sel The dataset after feature vector reduction is:

[0154]

[0155] In step 303, feature extraction is performed using 1DCNN, through parameter θ. numFilters and θ convlayers Convolution is performed to obtain the reduced feature dataset X exa And input it into the LSTM model, specifically:

[0156] 1DCNN contains m = 1, ..., M convolutional layers, with X1 as its input. sel X1 exists sel =X sel The specific process is as follows:

[0157] First, determine the kernel size as M. sel ×θ numFilters ;

[0158] Next, with a stride of s, the samples are convolved with the convolution kernel. The nth sample's feature count M is calculated after each convolution. sel Dimensionality reduction to θ numFilters The convolution process is as follows:

[0159]

[0160] In the formula, W k,1 b k,1 These are the convolution kernel and bias of the first layer, respectively. The input X for the first convolution layer sel The nth sample, This is the output of the nth sample from the first convolutional layer. The output is the result after the first convolutional layer has processed all samples. for:

[0161]

[0162] Similarly, the input to the m-th convolutional layer is the output of the (m-1)-th layer. The convolution process is as follows:

[0163]

[0164] In the formula, The input X for the m-th convolution layer exa The nth sample, W k,m b is the convolution kernel of the m-th layer. k,m The convolution bias for the m-th layer, The output of the nth sample in the m-th convolution layer, and the output after the m-th convolution layer. for:

[0165]

[0166] The output X of layer M exa for:

[0167]

[0168] In step 304, the parameter θ Epochs θ learningrate and θ dropout The prediction model is constructed, and the network parameters are updated according to the loss function during the modeling process, specifically as follows:

[0169] With parameter θ Epochs θ learningrate and θ dropout Construct an LSTM prediction model and reduce the feature dataset X. exa And input it into the LSTM prediction model, h represents the nth sample input. n-1 and Cn-1 h represents the output and state information of the LSTM hidden layer for the (n-1)th sample, respectively, and is used as the input for the next sample. n and C n The output and state information of the LSTM for the nth sample are respectively. The specific process during the initial sample calculation is as follows:

[0170] Define the empty set C n For memory units, forget gate f n Determine the memory unit C n The proportion of the previous sample output value that is output to the current input; the forget gate outputs F. n The calculation is as follows:

[0171]

[0172] F n =f n ⊙C n (16)

[0173] In the formula, U f and W f Corresponding to h respectively n-1 and The weight, b f For bias, σ(·) represents the sigmoid activation function, and ⊙ represents the Hadamard product, used to calculate candidate values ​​for the current input state. for:

[0174]

[0175] In the formula, U c and W c Corresponding to h respectively n-1 , The weight, b c For bias, tanh(·) denotes the tanh activation function, and the input gate i is calculated. n Used to control the storage of the nth sample input into memory unit C. n The ratio of input gate output I n for

[0176]

[0177]

[0178] In the formula, U i and W i Corresponding to h respectively n-1 and The weight, b i As a bias, the obtained value is stored in memory cell C. n ,get:

[0179] C n =F n +I n (20)

[0180] The first sample output C n The value is I n C n =I n ,F n =0, calculate the hidden layer output h of the nth sample. n Calculate the output gate o n for:

[0181]

[0182] h n =o n ⊙tanh(C n ) (twenty two)

[0183] In the formula, U o and W o Corresponding to h respectively n-1 and The weight, b o As a bias, calculate the output value of the nth sample. for:

[0184]

[0185] Among them, W out The weights correspond to the hidden layers. Similarly, when the remaining samples are input, the above process is repeated, and the output is... for:

[0186]

[0187] The loss function e′ is:

[0188]

[0189] In the formula, e′ represents The error between the actual data y and the feedback, y n It is the actual value of the nth sample. Based on f 1DCNN-LSTM The predicted value of the nth sample output by (·) is first updated by updating the convolutional layer parameters as follows:

[0190]

[0191]

[0192] in, The updated convolutional kernel and bias are respectively used. Next, the LSTM parameters are updated, and a hidden state C exists. n and h n ,definition and The corresponding calculation is as follows:

[0193]

[0194]

[0195] Calculate the weights and biases of the forget gate, input gate, and output gate. The learning formula for the output gate weights is as follows:

[0196]

[0197] The weights are updated as follows:

[0198]

[0199] In the formula, The updated output gate weights;

[0200] Calculate test performance metrics, calculate Ri for the test set. 2 As for particle fitness, the specific calculation process is as follows:

[0201]

[0202] As can be seen from the above formula, there are 6 hyperparameters involved, which influence each other and it is necessary to optimize them simultaneously.

[0203] In step 4, a PSO-based optimization process is performed to address the optimization problem, specifically as follows:

[0204] The modeling process is optimized based on the PSO algorithm, including population initialization, updating particle velocities and selecting decision variables, calculating fitness, updating individual particle optima, profiles, and sample pools, and obtaining the global optimum and model hyperparameters after reaching the required number of iterations. Population initialization is as follows: First, the number of particles P is... num Number of iterations N iter Update threshold N refresh The parameters, such as the upper and lower limits of the decision variables, are set, and then the result is generated by P. num A population of particles, with θ randomly initialized for each particle. p v p And calculate the particle's f(X), then initialize the particle's d. p Finally, with A, the particle's... With E P ;

[0205] The particle velocity is updated, and the decision variables are selected as follows: with the objective function of maximizing f(X), the particle's current best position is determined by the following formula: its velocity v p The update formula is:

[0206]

[0207] In the formula, w inertia The inertial weights that affect the particle search step size, where c is the learning factor. They follow a uniform distribution between [0,1]. For the nth dimension of particle p, the particle position θ p The update formula is:

[0208]

[0209] Fitness is calculated by calculating R0 on the test set. 2 As a particle fitness metric, the specific process is as follows:

[0210]

[0211] Update the individual particle optimality, file, and sample pool. The particle update formula is:

[0212]

[0213] In the formula, This represents the individual optimality of a particle. Furthermore, particle updates should not exceed the originally set range, i.e.:

[0214]

[0215] In the formula, pop represents the information of the particle, and popmin(site) and popmax(site) represent the minimum and maximum thresholds of the particle range at position site. If the individual particle is in the optimal iteration N refresh If no update is performed after this step, the learning sample pool will be updated. The strategy is to set the update probability of each dimension of the particle's learning sample as follows: During the update, first, two particles are randomly selected from the population. Then, the individual optimality of the two particles is compared. After the competition, the better individual optimality is selected as the new learning example, as follows:

[0216]

[0217] Learning probability of particle p The update is as follows:

[0218]

[0219] As particle sorting rank pAs the learning probability increases, the probability of updating the learned sample gradually increases within the range of 5% to 50%.

[0220] Calculating the global optimum: After initializing the population, the iterative optimization phase begins. First, the particle velocity v is updated. p Update the position θ of the particle's parameter decision variable. p After obtaining the hyperparameter values, feature reduction, model construction, and calculation of the evaluation index R are performed. 2 Using f(X) as the particle fitness value, update the individual optimality of the particle based on the fitness value, store the non-dominated solutions found in the population search into the archive, update the archive A, and calculate the individual optimal ranking rank of the particles. p And update its learning probability P c Therefore, for iteration N refresh After that, the particle whose individual optimal value has not been updated is updated with the learning sample. The particle with the largest f(X) value in the file is taken as the global optimal, and the optimal hyperparameter is obtained after decoding.

[0221] The method described in this invention was industrially verified. The computer hardware and software configuration for simulation verification was as follows: Windows 11 operating system, Matlab 2022; Intel i7 CPU, 16GB RAM. Specifically, this invention selects R... 2 The evaluation is based on three indicators: RMSE and MAE, using the following formula:

[0222]

[0223]

[0224]

[0225] In the formula, n test Indicates the nth test N samples, N test This represents the total number of samples in the test set. Indicates the nth test set test The actual value of each sample Indicates the nth test set test The predicted value for each sample, express The average value; the dataset of one embodiment of the present invention comes from a MSWI power plant in Beijing. The data within each minute is averaged, resulting in 720 groups, which contain 185 process variables and one output variable (CO concentration). The samples are divided into 4 equal parts at intervals. The first and second parts are used as training samples, the third part is used as validation samples, and the fourth part is used as test samples. The PSO algorithm parameters and model hyperparameters set according to the data characteristics are shown in Table 1.

[0226] Table 1. PSO algorithm and hyperparameter settings for the CO dataset.

[0227]

[0228] During the PSO iteration process, the changes in the values ​​of each hyperparameter are as follows: Figures 3a-3f As shown;

[0229] The optimal solution comes from the 24th particle. The relationship between the number of optimization iterations and the fitness function value in the CO dataset is as follows: Figure 4 As shown;

[0230] Depend on Figure 4 It can be seen that PSO tends to stabilize after the 17th iteration. 2 The value is 0.8193.

[0231] The parameters of the PSO-optimized MI-1DCNN-LSTM model are set as follows: θ MI The value is 0.6371, which corresponds to s = 1 and θ in 1DCNN. numFilters =9 and θ convlayers =7, which is θ in LSTM Epochs =371, θ learningrate =0.0075 and θ dropout =0.1. The comparison methods used were BP, RF, and LSTM, with the following parameters set: BP: 3000 iterations, convergence rate 0.0001, learning rate 0.01, and number of hidden layer neurons 371; RF: minimum number of leaf nodes 5 and number of decision trees 40; LSTM: 371 training iterations, learning rate 0.0075, and dropout rate 0.1. The experiments for each method were repeated 30 times. The statistical results of the evaluation metrics are shown in Table 2, and the test curves are shown in... Figures 5a-5c As shown;

[0232] Table 2 Comparison of experimental results for the CO dataset

[0233]

[0234] As shown in Table 2 and Figure 5: 1) The R values ​​of RF and LSTM in training, validation, and testing 2 1) MRSE and MAE scores are all better than BP, indicating that BP is difficult to apply to the CO dataset; 2) LSTM outperforms RF on both training and validation sets, but slightly underperforms RF on the test set. The variance of the LSTM model on the test set is 1.57E-04, while the variance of the RF model on the test set is 8.89E-04, indicating greater volatility and suggesting that LSTM has better predictive performance; 3) MI-1DCNN-LSTM has the best R-values ​​on both training, validation, and test sets. 2The statistical results of RMSE and MAE indicators show that MI-1DCNN-LSTM has good generalization performance and stability, and the model has the best prediction effect.

[0235] Experiments demonstrate the significant advantages of the proposed MI-1DCNN-LSTM model in constructing CO prediction models. Since different hyperparameters vary across different datasets, utilizing PSO for global hyperparameter optimization is reasonable in this invention.

[0236] Table 3. Meaning of symbols used in this invention

[0237]

[0238]

[0239]

[0240]

[0241]

[0242]

[0243] This invention provides a method for predicting CO emissions from the MSWI process. First, it employs particle design oriented towards reduced deep features and LSTM optimization, adaptively determining the feature selection threshold range based on the characteristics of the modeling data. Second, it inputs the deep features extracted using nonlinear features via super-one-dimensional convolution into an LSTM to construct a prediction model. The hyperparameters of the convolutional layers and LSTM are updated based on a loss function, with the model's generalization performance serving as the fitness function for the optimization algorithm. Finally, it uses a particle swarm optimization (PSO) algorithm to adaptively select features and hyperparameters based on data characteristics. The rationality and effectiveness of the method are verified on the MSWI process CO dataset, demonstrating its ability to predict CO emissions from the MSWI process based on reduced deep features and LSTM optimization.

[0244] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for predicting CO emissions from an MSWI process, characterized in that, Includes the following steps: Step 1: Construct an optimization problem for CO emission prediction based on reduced deep features and LSTM optimization; Step 2: For the optimization problem, conduct particle design oriented towards reducing deep features and LSTM optimization; Step 3: Design a fitness function for optimization problems that is oriented towards reducing deep features and LSTM optimization; Step 4: For the optimization problem, perform a PSO-based optimization process; In step 1, an optimization problem for CO emission prediction based on reduced deep features and LSTM optimization is constructed, specifically as follows: （1） In the formula, The model's performance index is used as the fitness function for optimization. For models based on 1DCNN and LSTM, Input the original sample dataset. for Normalized dataset, For MI feature selection The feature dataset, For the test set in Generalization performance is used as the fitness. , This is the feature dataset after feature extraction using 1DCNN. The total number of samples is [number]. indivual, The number of convolutional layers, total layer, for dimensionality for dimensionality for dimensionality and These represent the number of convolutional kernels and the number of convolutional layers, respectively.

2. The method for predicting CO emissions from the MSWI process according to claim 1, characterized in that, In step 2, for the optimization problem, particle design is carried out for reducing deep features and LSTM optimization, specifically as follows: The decision variables for designing particles are denoted as follows: ,Right now: （2） In the formula, For feature selection threshold based on MI, , and The MI value is calculated based on the modeling data, taking into account the number of iterations, learning rate, and dropout rate of the LSTM model. and As shown below: （3） （4） In the formula, and These are the maximum and average values ​​of MI, respectively. for The first in One characteristic, To normalize the CO concentration, various characteristics and The MI value between them is denoted as Based on the characteristics of the data, a suitable threshold range for modeling the data is obtained, i.e. Based on the settings get For each feature, the hyperparameters of subsequent 1DCNN feature extraction should satisfy the following relationship: LSTM hyperparameters are determined based on empirical terms. , and The upper and lower limits.

3. The method for predicting CO emissions from the MSWI process according to claim 2, characterized in that, In step 3, for the optimization problem, a fitness function is designed for reducing deep features and optimizing LSTM, specifically including the following steps: Step 301: Process the original dataset Normalization is performed to obtain a standard dataset. ; Step 302: Perform MI-based feature selection, and calculate the relationship between individual features in the process variables and CO concentration. The MI value between, and then based on get In 1 feature vector, obtain the feature dataset ; Step 303: Perform feature extraction using 1DCNN, through parameters... and Convolution is performed to obtain the reduced feature dataset And input it into the LSTM model; Step 304: Using parameters , and Build a prediction model and update the network parameters according to the loss function during the modeling process.

4. The method for predicting CO emissions from the MSWI process according to claim 3, characterized in that, In step 301, the original dataset is processed. Normalization is performed to obtain a standard dataset. Specifically: Will After normalization ,in, The calculation is shown in the following formula: （5） In the formula, for The i-th feature, and Let be the minimum and maximum values ​​of the i-th feature, respectively. Normalize the data to [0, 1]. , ,exist （6） Similarly, normalize y to .

5. The MSWI process CO emission prediction method according to claim 4, characterized in that, In step 302, feature selection based on MI is performed, and the relationship between individual features in the process variables and CO concentration is calculated. The MI value between, and then based on get In 1 feature vector, obtain the feature dataset Specifically: Calculate the relationship between a single characteristic and CO concentration in the process variables. To determine the MI value between sets, first, set the empty set. Calculate using the following formula The Middle The first sample eigenvalues and The MI values ​​between them are: （7） In the formula, For the first The first sample 1 eigenvalue, For the first The value of each sample For joint probability distribution, For the marginal probability distribution, calculate for: （8） Based on settings get In The dataset after feature vector reduction is: （9）。 6. The method for predicting CO emissions from the MSWI process according to claim 5, characterized in that, In step 303, feature extraction is performed using 1DCNN, through parameters. and Convolution is performed to obtain the reduced feature dataset And input it into the LSTM model, specifically: 1DCNN includes Each convolutional layer has an input of... ,exist The convolution process of the first layer is as follows: First, determine the kernel size as follows: ; Next, the samples are convolved with the convolution kernel. The convolution process is as follows: （10） In the formula, , These are the convolution kernel and bias of the first layer, respectively. The input for the first convolutional layer The One sample, For the first layer of convolution... The output of each sample, after the first convolutional layer completes the convolution of all samples. ,for: （11） Similarly, the output of the Mth layer for: （12）。 7. The method for predicting CO emissions from the MSWI process according to claim 6, characterized in that, In step 304, with parameters , and The prediction model is constructed, and the network parameters are updated according to the loss function during the modeling process, specifically as follows: Reduced feature dataset And input it into the LSTM prediction model, where, Indicates the first One sample input, and The first The output and state information of the hidden layer of the LSTM for each sample are used as the input for the next sample. and The first The specific process for calculating the LSTM output and state information for each sample during the initial sample computation is as follows: Define the empty set For memory units, forget gate Output The calculation is as follows: （13） （14） In the formula, and Corresponding to and The weight, For bias, This represents the sigmoid activation function. Represents the Hadamard product, used to compute candidate values ​​for the current input state. for: （15） In the formula, and Corresponding to , The weight, For bias, This represents the tanh activation function; Calculate input gate Output for: （16） （17） In the formula, and Corresponding to and The weight, As a bias, the obtained value is stored in the memory unit. ,get: （18） The first sample output The value is ,Right now Calculate the first Hidden layer output of each sample Calculate the output gate for: （19） （20） In the formula, and Corresponding to and The weight, For bias, calculate the first Output value of each sample for: （21） in, The weights correspond to the hidden layers. Similarly, when the remaining samples are input, the above process is repeated, and the output is... ,for: （22） loss function for: （23） In the formula, express Compared with actual data Feedback error, It is the first The actual value of each sample Based on The For each sample output prediction value, the convolutional layer parameters are first updated as follows: （24） （25） in, , The updated convolutional kernel and bias are then used. Next, the LSTM parameters are updated, and a hidden state exists. and ,definition and The calculation is as follows: （26） Calculate the weights and biases of the forget gate, input gate, and output gate. The learning formula for the output gate weights is as follows: （27） The weights are updated as follows: （28） In the formula, This refers to the updated output gate weights.

8. The method for predicting CO emissions from the MSWI process according to claim 7, characterized in that, In step 4, a PSO-based optimization process is performed to address the optimization problem, specifically as follows: The modeling process is optimized based on the PSO algorithm, which includes population initialization, updating particle velocity and selecting decision variables, calculating fitness, updating individual particle optimality, profile and sample pool, and obtaining global optimum and model hyperparameters after reaching the number of iterations. Population initialization is as follows: First, the number of particles is... Number of iterations Update threshold The parameters, such as the upper and lower limits of the decision variables, are set, and then a process is generated. A population of particles, with particles randomly initialized. , And calculating particles Next, initialize the particles. and Finally, calculate the particle's and ; Update particle velocity and select decision variables, in order to Maximize the speed of the objective function. The update formula is: （29） In the formula, The inertial weights that affect the particle search step size As a learning factor, obey Uniform distribution between For particles No. Dimensional learning examples, particle positions The update formula is: （30） Fitness is calculated by calculating R0 on the test set. 2 As a particle fitness metric, the specific process is as follows: （31） The particle update formula is as follows: (Update individual particle optimality, file, and sample pool.) （32） In the formula, In the individual optimal formula for a particle... Information representing particles, and Represent the minimum and maximum thresholds for the particle range at position site, if the individual particle is in the optimal iteration If no update is performed after this step, the learning sample pool will be updated. The strategy is to set the update probability of each dimension of the particle's learning sample as follows: During the update, two particles are randomly selected from the population. Their individual optimal values ​​are then compared. After the competition, the better individual optimal value is selected as the new learning example. （33） particle learning probability The update is as follows: （34） With particle sorting As the learning probability increases, the probability of updating the learned sample gradually increases within the range of 5% to 50%. Calculating the global optimum: After initializing the population, the iterative optimization phase begins. First, the particle velocities are updated. Update the position of the particle's parameter decision variables. After obtaining the hyperparameter values, feature reduction, model construction, and calculation of the evaluation index R are performed. 2 Used as particle fitness value Based on the fitness value, the optimal value of each particle is updated, and the non-dominated solutions found by the population search are stored in the archive, which is then updated. Calculate the individual optimal ranking of particles and update its learning probability. Therefore, the iteration The next step is to update the learning examples for particles whose optimal individual performance has not yet been updated, and to update the archive. The particle with the largest value is taken as the global optimum, and the optimal hyperparameters are obtained after decoding.

Images