In-furnace flame combustion line position modeling method in urban solid waste incineration process

Abstract

The invention belongs to the technical field of urban solid waste incineration, and provides an urban solid waste incineration process in-furnace flame combustion line position modeling method which comprises the steps of process video and data acquisition and combustion line modeling. The in-furnace flame combustion line position modeling model training process comprises combustion line dynamic quantification processing, filling processing and IT2FFR-LSTM main and complementary model training; according to the method, dynamic quantization processing is performed on the combustion line through GAN-SCNN, grid position features are enriched, and a corresponding template is provided for a subsequent control strategy; data filling processing is performed by using RF and PLS, so that insufficient adaptability of simple pre-filling to a complex missing scene is avoided, more complete original data features are reserved for a subsequent filling algorithm, and the overall data quality is improved; and through the IT2FFR-LSTM main compensation model, the preliminary prediction result is corrected, and the modeling effect is improved.

Description

Technical Field

[0001] This invention relates to the field of urban solid waste incineration technology, and in particular to a method for modeling the position of the flame combustion line in the furnace during urban solid waste incineration. Background Technology

[0002] Municipal solid waste incineration (MSWI) is a widely used treatment method worldwide, offering advantages such as harmlessness, volume reduction, and resource recovery. Currently, MSWI technology in developing countries faces numerous challenges, the most prominent being unstable combustion due to manual, experience-based operation, leading to significant fluctuations in pollutant emissions. Furthermore, unstable combustion can cause malfunctions such as coking, ash accumulation, and corrosion within the furnace, and in severe cases, even furnace explosions. Therefore, maintaining stable combustion is crucial for ensuring efficient MSWI operation and compliance with emission standards. Achieving stable combustion in MSWI requires timely adjustments to manipulated variables such as feed, air, and water based on changes in controlled variables. Furnace temperature, flue gas oxygen content, steam flow rate, and flame combustion position are commonly controlled variables. While modeling and control studies for the first three controlled variables have been reported, research on modeling and control of the flame combustion line position is yet to be published.

[0003] Unlike the other three controlled variables, which can be monitored in real time, the flame burn line position is not perceived or detected in real time. It is mainly quantified by manual observation and experience in the minds of domain experts. Therefore, modeling the flame burn line position requires solving three problems in sequence: obtaining the truth value, constructing the modeling dataset, and building the model. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method for modeling the position of the flame combustion line in the furnace during urban solid waste incineration, thereby solving the problem that existing methods rely on manual experience.

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

[0006] A method for modeling the flame combustion line position in an urban solid waste incineration furnace includes:

[0007] Collect video of the flame to be processed and data of the processing process;

[0008] The flame video to be processed and the process data to be processed are input into the pre-trained furnace flame combustion line position modeling model to perform combustion line modeling, thereby obtaining the furnace flame combustion line position modeling result; the training process of the furnace flame combustion line position modeling model includes:

[0009] GAN-SCNN was used to perform dynamic quantization of the combustion lines on the pre-collected flame combustion images to obtain the quantized values ​​of the flame combustion line positions.

[0010] Set up operation variables, and use RF and PLS to fill in the operation variables and the quantization values ​​of the flame combustion line position corresponding to the flame combustion image to obtain the modeling dataset;

[0011] The IT2FFR-LSTM master complement model is trained using the modeling dataset to obtain the modeling model of the flame combustion line position in the furnace.

[0012] The present invention discloses the following technical effects:

[0013] This invention provides a method for modeling the position of the flame combustion line in the furnace during urban solid waste incineration. By dynamically quantizing the combustion line using GAN-SCNN, it solves the problem that traditional image generation algorithms struggle to supplement unique flame features in sparse images when data is limited, thus enriching the grid position features and providing corresponding templates for subsequent control strategies. By utilizing RF and PLS for data filling, it addresses the issue of missing data at certain times due to sensor corrosion from high-temperature flue gas and dust adhesion in MSWI field monitoring environments, preserving complete original data features for subsequent filling algorithms. Finally, by employing the IT2FFR-LSTM master complement model, it solves the problem of large prediction errors in traditional models, enabling the correction of preliminary prediction results. Attached Figure Description

[0014] 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.

[0015] Figure 1 This is a schematic diagram of the process for modeling the position of the flame combustion line in the furnace during the urban solid waste incineration process, provided in an embodiment of the present invention.

[0016] Figure 2 This is a diagram showing the correspondence between the camera imaging position and the furnace shutdown photo provided in an embodiment of the present invention;

[0017] Figure 3 A modeling strategy diagram provided for embodiments of the present invention;

[0018] Figure 4 A partial flame image of the left grate provided in an embodiment of the present invention;

[0019] Figure 5 A schematic diagram illustrating the flame image quantization process provided in an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] The purpose of this invention is to provide a method for modeling the position of the flame combustion line in the furnace during urban solid waste incineration, thereby solving the problem that existing methods rely on manual experience.

[0022] 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.

[0023] Figure 1 This is a schematic diagram of the modeling process for the flame combustion line position in the furnace during the urban solid waste incineration process, provided in an embodiment of the present invention. Figure 1 As shown, this invention provides a method for modeling the flame combustion line position in an urban solid waste incineration furnace, including:

[0024] Step 100: Collect the video of the flame to be processed and the data of the processing process;

[0025] Step 200: Input the flame video to be processed and the process data to be processed into the pre-trained furnace flame combustion line position modeling model to perform combustion line modeling, and obtain the furnace flame combustion line position modeling result; the training process of the furnace flame combustion line position modeling model includes:

[0026] Step 201: Use GAN-SCNN to perform dynamic quantization processing on the pre-collected flame combustion images to obtain the quantized values ​​of the flame combustion line positions;

[0027] Step 202: Set operation variables, and use RF and PLS to fill in the operation variables and the quantization values ​​of the flame burning line position corresponding to the flame burning image to obtain the modeling dataset;

[0028] Step 203: Train the IT2FFR-LSTM master complement model using the modeling dataset to obtain the modeling model of the flame combustion line position in the furnace.

[0029] Specifically, the MSWI process analysis for flame combustion line position modeling describes the principle of obtaining the true value of the flame combustion line position based on the combustion flame. The description uses the in-furnace combustion flame imaging of an 800t / d stepped (inclined + horizontal) forward-push mechanical grate furnace in a certain MSWI power plant as an example. The correspondence between the camera imaging position and the shutdown photo is shown in the figure below. Figure 2As shown. The camera video channel resolution is 720. 576, meaning the image width is 576 pixels. Since the drying grate could not be fully imaged by the camera, the pixel points corresponding to the actual positions within the furnace were calculated using a 3D spatial-to-pixel mapping relationship. The calculation results are shown in Table 1.

[0030] Table 1

[0031]

[0032] During incineration, considering the influence of the material layer thickness, the imaging ratio at the front of the drying and combustion sections needs to be corrected by 5%. Table 1 shows that the normal range for the flame burn line position is 51% to 73.6%. Clearly, an image library of images from different combustion states is needed to support the acquisition of the true flame line value.

[0033] Furthermore, the analysis of factors influencing the flame combustion line position reveals that, in order to achieve stable combustion, meet emission standards, and improve thermal efficiency in the MSW process, it is necessary to continuously monitor and appropriately adjust many process variables. Due to the strong coupling between process variables in the MSW system, relying solely on statistical correlation often fails to identify variable combinations meaningful from a control perspective. Therefore, a guided approach based on domain expert rules was adopted, rooted in actual industrial control requirements and field operational knowledge. The manipulated variables and their value ranges under the selected baseline operating conditions are shown in Table 2 below.

[0034] Table 2

[0035]

[0036] Specifically, the modeling strategy. This embodiment proposes a control-oriented flame combustion line position modeling strategy, which consists of three parts: a flame combustion line position dynamic quantization module based on flame combustion images, a dataset acquisition module based on data imputation and expert knowledge, and a fireline controlled object model module based on the IT2FFR-LSTM master complement mechanism. Figure 3 As shown. Figure 3 middle, Represents the original flame video, This represents a single frame of a flame image captured on a minute-by-minute scale. A quantized value representing the position of the flame burn line. Represents the original process data. This represents the preprocessed process data. Key manipulated variables representing choices, Represents the input of manipulated variables. Represents the actual output of the front line. This represents the prediction output of the IT2FFR master model. Represents the error of the main model , This represents the predicted output of the LSTM compensation model. This represents the predicted output of the IT2FFR-LSTM master complement model. Figure 3 The functions of each module are as follows:

[0037] 1) Dynamic quantization module for flame burning line position based on flame burning image: Through flame image preprocessing and flame burning line position quantization processing, a quantized value that can reflect the dynamic position of the flame burning line is obtained;

[0038] 2) Dataset acquisition module based on data imputation and expert knowledge: The key manipulated variable dataset is obtained through process data matching preprocessing, key manipulated variable selection, and prediction imputation of missing values;

[0039] 3) Firefighting Controlled Object Model Module Based on IT2FFR-LSTM Compensation Mechanism: First, the main model IT2FFR captures the uncertainty of the input data and outputs preliminary prediction results. Then, the compensation model LSTM learns the prediction error of the main model and corrects its predictions. Finally, the results of the main model and the LSTM compensation are superimposed as the final output, thereby constructing the firefighting controlled object model.

[0040] Furthermore, the algorithm is implemented. The dynamic quantization module for flame line positions based on flame combustion images comprises two parts: offline construction of a complete image library based on GAN and dynamic quantization of the combustion line based on SCNN.

[0041] 1) Offline Construction Submodule of Complete Image Library Based on GAN: Data-driven CLQ algorithms require data consistent with the global distribution. However, with limited data, traditional image generation algorithms struggle to supplement unique flame features in sparse images, making it more difficult to capture missing image features. Therefore, this embodiment employs image processing, Deep Convolutional Generative Adversarial Network (DCGAN), Cycle Consistent Generative Adversarial Network (CycleGAN) techniques and mechanistic knowledge to process furnace flame images, constructing a complete flame image library containing normal, abnormal, and highly abnormal combustion modes. This library is enriched with raster location features, a large number of unlabeled flame images, and mechanism-based pseudo-labels.

[0042] 2) SCNN-based Combustion Line Dynamic Quantization Submodule: This embodiment aims to improve quantization accuracy by incorporating a wider range of flame characteristics and providing more flame information for subsequent control purposes. Therefore, this embodiment employs a Spatial Convolutional Neural Network (SCNN) matching method for quantization. This strategy not only generates quantized values ​​but also provides corresponding templates. These matched templates can be directly linked to the control strategy, thereby realizing an "end-to-end" control method. First, the SCNN is trained based on a complete flame image library; then, combustion line features are extracted, and the corresponding template sub-library is loaded; finally, online CLQ is implemented based on SCNN multi-scale feature similarity matching.

[0043] Specifically, the flame image preprocessing submodule. First, after reading the complete flame video, it parses the video's basic parameters (frame rate). Total duration Total Frames ) and define frame capture rules (setting "the same moment every minute" to a fixed number of seconds). For example, at the 10th second of every minute, Accordingly, the sequence of target frame timestamps is as follows:

[0044]

[0045] in, Non-negative integers ( =0 corresponds to the target time in the 1st minute. =1 corresponds to the target time in the 2nd minute). Set a fixed number of frames per minute (e.g.) =10, which is the 10th second of every minute).

[0046] Next, the frame index corresponding to each target time point is calculated based on the frame rate, and the image of that frame is located and extracted:

[0047]

[0048] in, For the first The index (integer) of the target frame. This is a rounding function (ensuring the frame index is an integer to avoid locating invalid frames).

[0049] Finally, the extracted frames are format-normalized to obtain a set of single-frame flame images that match the process variables.

[0050] Further, a flame combustion line position quantization submodule is implemented. A complete image library is constructed offline based on GAN: First, a combustion line edge feature extraction algorithm is designed to extract the combustion line feature map, which includes steps such as inverse binarization, median filtering, dilation, and combustion line edge feature calculation. This process is shown in the following equation:

[0051]

[0052] In the formula, This represents the input flame image; Indicates inverse binarization, The threshold for inverse binarization; Indicates median filtering; This indicates that 40 expansion operations are performed, with the operator size being (1, 5), and the purpose is to eliminate small, scattered flames from a lateral angle. The operator calculates the edge features of the combustion line, that is, it uses the vertical gradient from bottom to top to represent the position of the transverse combustion line.

[0053] Then, a combustion line position calibration algorithm is used to extract combustion line feature values. Morphological processing algorithms are then used to process the flame image to obtain combustion line edge information images. Then, calculate The mean value of the ordinate of the white pixels representing the combustion line As shown in the following formula:

[0054]

[0055] In the formula, and They represent The resolution refers to the width and height of the image captured by the camera. In this embodiment, the values ​​are 720 and 576, respectively. This represents the pixel in the i-th row and j-th column of the input image. This indicates the lower limit of the burning area in pixels. `count` indicates the range of pixels within the [...] pixel width. , ], length [0, Pixels within the region satisfy The number of.

[0056] Next, calculate Variance of the ordinate of the white pixel As shown in the following formula:

[0057]

[0058] Finally, when Less than At that time, the combustion line calibration result res was ,when Greater than When res is marked as 0, it indicates an unconventional combustion line, as shown in the following formula:

[0059]

[0060] Furthermore, dynamic quantization of combustion lines based on SCNN: First, the feature map of the combustion lines is obtained based on morphological processing algorithms, as shown in the following equation:

[0061]

[0062] in, For combustion line calibration algorithm; Real-time flame images;

[0063] Next, combustion line features are extracted based on the combustion line position calibration algorithm, as shown in the following formula:

[0064]

[0065] in, for The mean value of the combustion line; for The variance of the combustion line; when Greater than When, output A value of 0 indicates that the combustion line in the real-time flame image does not exist; when Less than At that time, a similarity measurement module based on SCNN is used, which includes functions such as template library loading, Siamese network, and template reloading.

[0066] according to Load the corresponding template library according to the correspondence with Table 1. As shown in the following formula:

[0067]

[0068] In the formula, This indicates the i-th template in the corresponding template sub-library that has been loaded. A library of extreme and unusual flame image templates; A template library containing both real and generated images of flames with forward and extreme forward combustion lines; A template library containing a set of normal flame images with realistic combustion lines; A template library containing both real and generated images of flames with backward-shifted combustion lines; Medium template Describing features , , This indicates the number of templates in the current template sub-library.

[0069] Using SCNN metrics and The similarity is calculated, and the maximum value is taken, as shown in the following formula:

[0070]

[0071]

[0072] in, For Siamese network-based metrics and Similarity; For the Siamese network algorithm; It is a fully connected layer in a twin network; This refers to VGG16 layer in a twin network; Indicates the use of SCNN metrics middle and The maximum similarity;

[0073] After calculating the similarity, Greater than the similarity threshold The mean is used as the matching template.

[0074] Based on similarity and template feature description, setting The value is 0.99. A weighted algorithm is used to solve the quantization value of the burning lines in the current image. ,when All less than Obtained using classical algorithms As the quantification value of the combustion line, the process is shown in the following formula:

[0075]

[0076]

[0077] in, These are weight parameters; for Number of templates.

[0078] Specifically, the dataset acquisition module is based on data imputation and expert knowledge. The process data matching and preprocessing submodule, in order to construct an experimental raw dataset that truly reflects the combustion process characteristics of the MSWI system, first employs a "time-dimensional random uniform sampling" strategy for the continuous operation monitoring data of the MSWI system: a fixed second-level time point is randomly selected from 0 to 59 seconds per minute, serving as a unified time benchmark. Within the entire monitoring cycle, multi-dimensional process data corresponding to that second-level time point and the quantized value of the flame combustion line position at that time are extracted every minute. These are then integrated according to the association structure of "timestamp - process data - quantized value of flame combustion line position" to form the experimental raw dataset. Considering that sensors are susceptible to corrosion from high-temperature flue gas and dust adhesion in the MSWI field monitoring environment, resulting in missing data at certain times, this study further adopts a "scenario-based stepped filling" strategy to preprocess missing data: For "single isolated missing" scenarios (i.e., data is missing at a certain time, but the data at the previous and next adjacent times are complete and valid), a weighted average method of the data at previous and next times is used for pre-filling based on the continuity and recency effect of time series data; For other complex missing scenarios (including missing data at multiple consecutive times, missing times at the beginning and end of the dataset, or isolated missing data at both the previous and next adjacent times), pre-filling is not performed for the time being, and it is reserved for accurate filling later in combination with MSWI process mechanism or data-driven model. This avoids the insufficient adaptability of simple pre-filling to complex missing scenarios, and also preserves more complete original data features for subsequent filling algorithms, thereby improving the overall data quality.

[0079] Optionally, in the key manipulated variable selection submodule, in order to construct a control-oriented MSWI process flame burnline position model, this embodiment determines the key manipulated variables related to the flame burnline position based on domain expert knowledge and the actual process mechanism on site.

[0080] Preferably, a missing value filling submodule is used. Missing values ​​in the process data and flame combustion line positions present during the acquisition of process data and the quantization of the actual video flame images of the incineration plant are first extracted using mutual information, and then filled using RF (Random Forest) or PLS (Partial Least Squares) algorithms, according to the following rules:

[0081]

[0082] in, To predict the output feature vector, For RF padding algorithm, For PLS filling algorithm, The number of samples with missing features. The number of samples with valid features.

[0083] The method trains the valid process data and quantized values ​​along with their corresponding multiple related features. Then, it uses the remaining missing process data and the corresponding multiple related features as input to predict the valid process data and quantized values. The specific process is as follows: First, the features with missing values ​​are used as the "target variable," and other features without missing values ​​in the dataset are used as "input features." A training set is constructed by selecting all samples with no missing target variables (ensuring the integrity of the training data). Next, an RF / PLS model is built. Finally, the trained model is used to predict the samples with missing target variables, and the prediction results are used as imputation values. The advantage of this method is its ability to capture complex nonlinear relationships between features and its strong robustness to outliers. A simplified description of the imputation algorithm is shown below.

[0084] 1) RF-based padding:

[0085] The core of Random Forest (RF) imputation of missing values ​​relies on its "multi-decision tree ensemble learning" characteristic. It achieves accurate imputation by mining the correlation between complete features and features containing missing values ​​in the dataset, making it particularly suitable for non-linear, multi-feature coupled data scenarios. First, let the input feature matrix and output vector be denoted as... and It will be generated via Bootstrap The process of training a subset is described as follows:

[0086]

[0087] in, It is the first The second selected training subset; Indicates the first The number of input features contained in each training subset.

[0088] Then, duplicate samples in the training subset are removed and the result is taken as the first... input feature vectors As a dividing variable, with the first The value corresponding to each sample Use these as dividing points for cutting.

[0089] Next, the optimal splitting variable and splitting point are obtained by traversal until the number of leaf node samples is less than an empirically set threshold. .

[0090] Finally, the input feature space is divided into Each region was marked as Therefore, the RF-based prediction model is constructed as follows:

[0091]

[0092]

[0093] in, It is a region Inner The training subset of the first One true value; For indicator functions, when If it exists, the function value is 1; otherwise, it is 0.

[0094] 2) PLS-based filling:

[0095] The PLS model employs a dual strategy of "extracting principal components of features + maximizing the correlation between features and the target variable." It is suitable for imputing missing values ​​in data with multicollinearity and high feature dimensionality, and is particularly well-suited for data scenarios with linear or weakly nonlinear associations. It can retain key correlation information while reducing dimensionality, avoiding prediction bias caused by feature redundancy. Its regression coefficients... The data is divided into two rows, corresponding to the "intercept coefficient" and the "feature regression coefficient" respectively:

[0096]

[0097] in, These are the PLS regression coefficients; Indicates original missing features The column mean, Represents the original input variables The column mean, Represents the PLS weight matrix. express Transpose of the load matrix.

[0098] The predicted value is calculated using the following formula:

[0099]

[0100] in, Original missing features Prediction matrix, Represents the original input variable matrix. It is an n-dimensional vector of all 1s.

[0101] Specifically, the Fireline Controlled Object Model Module is based on the IT2FFR-LSTM master-complement mechanism. The Interval Type II Fuzzy Forest Regression-Long Short-Term Memory Network (IT2FFR-LSTM) with master-complement mechanism proposed in this embodiment is characterized by: integrating multiple Interval Type II Fuzzy Decision Trees (IT2FDT) through the master model (IT2FFR) to capture the uncertainty of the input data and output preliminary prediction results; learning the prediction error of the master model through the compensation model (LSTM) to correct it; and summing the prediction results of the master and compensation models as the final output.

[0102] Further, the IT2FFR main model: First, data preprocessing is performed. Here, the total sample... Recorded as:

[0103]

[0104] in, Represents the number of samples. This represents the number of features. Also, let the input dataset be denoted as... The output dataset is To avoid the impact of differences in feature magnitude on fuzzy feature calculation and LSTM gradient update, zero-mean-unit variance normalization is applied:

[0105]

[0106]

[0107] in, To normalize the input dataset; To normalize the output dataset; for The mean and standard deviation, for The mean and standard deviation.

[0108] Next, Bootstrap sampling and feature selection are used to generate... A subset of Bootstrap training data is shown below:

[0109]

[0110]

[0111] in, For random sampling index with replacement, ( (The number of trees). Each subset is randomly selected. The process of each feature is represented by the following formula:

[0112]

[0113] in, For indexes of random features without replacement, ; The number of features randomly selected for each tree; For the first tree from Selected A subset of features; For the first in the forest The index of the tree; , For the first Input / output feature matrices of the Bootstrap sampling training subset of the tree.

[0114] Then, the interval type II fuzzy regression tree is constructed. For the first... The first tree Split the nodes according to the minimum MSE, thus solving the following optimization problem:

[0115]

[0116] in, This indicates obtaining the value that minimizes the objective function. function, For the first The first of the trees Index of each node; For the first The first of the trees The optimal splitting threshold for each node (determined by the minimum MSE criterion). The number of node samples. , These are the true value vectors for the left and right subsets, respectively. These represent the number of samples in the left and right subsets after the split. The leaf node The antecedent of the rule is an interval type II fuzzy set. The upper and lower bounds of the membership degree are as follows:

[0117]

[0118]

[0119] in, The leaf node An index of fuzzy rules; For interval type II fuzzy sets The upper bound of the membership function; For interval type II fuzzy sets The lower bound membership function; Centered on, For width, As an uncertainty factor, For the first One input variable, This refers to the index of the feature / variable. For the k-th tree in the forest, the interval prediction value is output through interval type II fuzzy rule inference within its structure. :

[0120]

[0121]

[0122] in, Let be the number of fuzzy rules for this tree. The number of features selected for this tree; Output interval predicted values ​​for interval type II fuzzy rule inference. Normalized activation strength of the rule As shown in the following formula:

[0123]

[0124] in, For the intercept interval of the rule consequent, The range represents the characteristic coefficients.

[0125] Next, the optimal weights are solved using ridge regression. (That is, the fusion weight vector) is shown below:

[0126]

[0127] in, Output a matrix for the tree. Let be the regularization coefficient. The analytical solution is:

[0128]

[0129] in, It is the identity matrix;

[0130] Furthermore, regarding The interval prediction results for each tree are obtained using the optimal weights learned during the training phase. Perform weighted fusion and use interval coefficients Integrate the upper and lower bounds as shown below:

[0131]

[0132] in, Interval coefficients;

[0133] Finally, the prediction output of the main model is denoted as follows:

[0134]

[0135] in, The final normalized prediction output of the main model, The standard deviation of the output variable; This is the mean of the output variable.

[0136] Specifically, the LSTM compensation model. The compensation model learns the error of the main model. To correct the prediction results, the modeling dataset is denoted as follows: The specific description of the LSTM network structure is as follows:

[0137] 1) LSTM gating mechanism:

[0138] First, the dimension of the input layer is the same as the number of input features of the main model.

[0139] Then, the LSTM layer handles timing dependencies through a gating mechanism. It consists of the following parts:

[0140] Forget Gate: Controls the preservation of historical cell states.

[0141]

[0142] Input gate and candidate cell status:

[0143]

[0144]

[0145] Cell status update:

[0146]

[0147] Output gates and hidden states:

[0148]

[0149]

[0150] Output layer: Outputs error predictions through a fully connected layer.

[0151]

[0152] in, For the sigmoid function, Hyperbolic tangent activation function, As input features, For element-wise product, This is the weight matrix for the gating mechanism. This is the bias term for the gating mechanism. for Hide your status at all times. In cellular state, This is the weight matrix of the output layer (which maps the hidden states to error predictions). For the bias term of the output layer; For LSTM in the first The forget gate outputs at any given moment; For LSTM in the first Input gate output at any given time; For LSTM in the first The state of candidate cells at any given time; For LSTM in the first Output gate at any given time.

[0153] 2) Loss Function and Optimization:

[0154] Using mean squared error (MSE) as the loss function :

[0155]

[0156] The Adam optimizer is used to update the weights through backpropagation. :

[0157]

[0158] Wherein, the learning rate is Iterate until the loss function converges or the maximum number of epochs is reached.

[0159] The total output of IT2FFR-LSTM is the superposition of the primary model prediction and LSTM error compensation:

[0160]

[0161] Preferably, the experimental results and analysis, data description: The data in this embodiment comes from flame videos and process data of a certain MSWI power plant during a certain period of 5 days in 2019, and the dataset contains a total of 780 samples. Single frame images are extracted from the flame videos on a minute time scale and quantized to obtain the quantized value of the flame combustion line at each moment. Finally, these values ​​are matched one-to-one with the process data at the same moment to form a complete multimodal dataset. Primary air temperature, primary air volume, primary air pressure, drying section air volume, combustion section 1 air volume, combustion section 2 air volume, burnout section exhaust air volume, secondary air temperature, secondary air volume, feeding rate, drying section grate speed, combustion section 1 grate speed, combustion section 2 grate speed, boiler feedwater, and urea are used as manipulated variables to model the quantized values ​​of the combustion lines of the left and right grates using the IT2FFR-LSTM model. The evaluation indicators used in this embodiment are root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The calculation formula is as follows:

[0162]

[0163]

[0164]

[0165] in, Represents the actual value. Represents the predicted value. This represents the average of the actual values.

[0166] Specifically, the parameter settings and performance evaluation index calculation results for the modeling process are shown in Table 3.

[0167] Table 3

[0168]

[0169] Looking at the various metrics of the dataset, the training set performs best overall, with the smallest RMSE (L = 0.0205, R = 0.0277) and MAE (L = 0.0163, R = 0.0214). 2 The highest values ​​(L = 0.8613, R = 0.8479) indicate that the model fits the training data well. As the data transitions from the training set to the validation set and then to the test set, both RMSE and MAE generally increase, while R... 2 The results showed a downward trend, with the R (0.6918) on the validation set decreasing significantly. The L (0.7344) and R (0.7217) on the test set were relatively similar, indicating that the model's generalization ability on unseen data was somewhat weakened, but it still maintained a certain explanatory power.

[0170] Furthermore, method comparison. To verify the effectiveness of the method proposed in this embodiment, IT2FBLS, IT2FNN, BPNN, LRDT, and FFR were used to model the quantized values ​​of the combustion line in the MSWI process, and the modeling results were compared with those of the IT2FFR-LSTM method in this embodiment. The IT2FBLS model parameters were set as follows: (left side of the grate: number of fuzzy components in the subsystem) The number of rules in the subsystem is 70. The initial number of enhanced nodes is 20. The scaling factor is 50. The enhancement layer scaling factor is 0.5. The regularization coefficient is 0.7. 1e-5; Right side of the grate: Fuzzy quantity of subsystems The number of rules in the subsystem is 70. The initial number of enhanced nodes is 20. The scaling factor is 90. The enhancement layer scaling factor is 0.5. The regularization coefficient is 0.8. The IT2FNN model parameters are set as follows: upper and lower bound scaling factor of 0.3, learning rate of 0.01, number of rules of 10, and number of iterations of 800; the BPNN model parameters are set as follows: number of hidden layer neurons of 31, maximum number of convergence iterations of 1500, convergence error of 0.01, and learning rate of 0.01; the LRDT model parameters are set as follows: minimum number of samples of 30, number of features of 4, and regularization coefficient of 0.3; the FFR model parameters are set as follows: learning rate of 0.1, intermediate variable (controlling model complexity) of 8, batch number of 3, least squares regularization coefficient of 0.5, minimum number of samples in decision tree of 30, number of features selected per tree of 4, and number of trees in forest of 85. The modeling comparison results are shown in Table 4.

[0171] Table 4

[0172]

[0173] Table 4 shows that the performance of different methods varies across the training, validation, and test sets, with the IT2FFR-LSTM method showing the best overall performance. On the training set, the FFR method achieves the best results in various metrics (L: RMSE=0.0195, MAE=0.0153, R...). 2 =0.8743; R: RMSE=0.0265, MAE=0.0208, R 2 =0.8602) is the best, followed by IT2FFR-LSTM and IT2FNN; on the validation set, IT2FFR-LSTM has the best RMSE, MAE and R on both the left and right sides. 2All are optimal, with good overall performance; on the test set, IT2FFR-LSTM is on the left (RMSE=0.0263, MAE=0.0200, R... 2 =0.7344) and the right side (RMSE=0.0367, MAE=0.0275, R 2 The R² value (=0.7217) significantly outperformed other methods in all metrics, demonstrating stronger generalization ability. In contrast, methods like IT2FNN and LRDT performed relatively weakly on the test set, especially R² on the right side. 2 The generally low values ​​indicate that it is not well adapted to new data.

[0174] The beneficial effects of this invention are as follows:

[0175] This invention uses GAN-SCNN to dynamically quantize the combustion line, enriching the grid position features and providing a corresponding template for subsequent control strategies. By using RF and PLS for data imputation, it avoids the insufficient adaptability of simple pre-filling to complex missing scenarios, preserves more complete original data features for subsequent imputation algorithms, and improves the overall data quality. The IT2FFR-LSTM master imputation model is used to correct the preliminary prediction results and improve the modeling effect.

[0176] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0177] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present 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 the present invention.

Claims

1. A method for modeling the position of the flame combustion line in an urban solid waste incineration furnace, characterized in that, include: Collect video of the flame to be processed and data of the processing process; The flame video to be processed and the process data to be processed are input into the pre-trained in-furnace flame combustion line position modeling model to perform combustion line modeling, and the in-furnace flame combustion line position modeling result is obtained. The training process of the furnace flame combustion line position modeling model includes: GAN-SCNN was used to perform dynamic quantization of the combustion lines on the pre-collected flame combustion images to obtain the quantized values ​​of the flame combustion line positions. Set up operation variables, and use RF and PLS to fill in the operation variables and the quantization values ​​of the flame combustion line position corresponding to the flame combustion image to obtain the modeling dataset; The IT2FFR-LSTM master complement model is trained using the modeling dataset to obtain the modeling model of the flame combustion line position in the furnace.

2. The method for modeling the flame combustion line position in an urban solid waste incineration furnace according to claim 1, characterized in that, The pre-collected flame combustion images were dynamically quantized using GAN-SCNN to obtain quantized values ​​of the flame combustion line positions, including: Frame extraction is performed on the flame combustion images to obtain a set of single-frame flame images; The combustion lines of the single-frame flame image set were calibrated using GAN to obtain the combustion line calibration results. The combustion line calibration results are dynamically quantized using SCNN to obtain the quantized value of the flame combustion line position.

3. The method for modeling the position of the flame combustion line in a furnace during urban solid waste incineration according to claim 1, characterized in that, By setting operational variables and using RF and PLS to fill in the operational variables and the quantized values ​​of the flame combustion line position corresponding to the flame combustion image, a modeling dataset is obtained, including: A filling algorithm is selected according to a preset filling rule; the filling algorithm includes either the RF prediction model or the PLS prediction model; the expression of the filling rule is: ;in, To predict the output feature vector; For RF padding algorithm; PLS filling algorithm; The number of samples with missing features; The number of samples without missing features; An initial dataset is constructed based on the operational variables and the quantized values ​​of the flame combustion line position. Features with missing values ​​in the initial dataset are set as target variables, and features without missing values ​​in the initial dataset are set as input features, thus obtaining the basic dataset. When the padding algorithm is RF, several training subsets are generated using Bootstrap based on the base dataset; The training subset is traversed through the segmentation variables and segmentation points until the number of leaf node samples is less than an empirically set threshold. Then, the input feature space is divided into several regions, and the RF prediction model is constructed based on the regions of the input feature space. The base dataset is populated using the RF prediction model to obtain the first populated dataset; The base dataset is populated using the PLS prediction model to obtain a second populated dataset; the expression of the PLS prediction model is: ;in, ; The input feature matrix; These are the PLS regression coefficients; A vector of all 1s; Original missing features The column mean; For the original input variables The column mean; This is the PLS weight matrix; for Transpose of the load matrix; The first filled dataset and the second filled dataset are integrated to obtain the modeling dataset.

4. The method for modeling the flame combustion line position in an urban solid waste incineration furnace according to claim 1, characterized in that, The IT2FFR-LSTM master complement model is trained using the aforementioned modeling dataset to obtain the furnace flame combustion line position modeling model, including: The modeling dataset is normalized to zero mean and unit variance to obtain a preprocessed dataset; Bootstrap sampling and random feature selection are performed on the preprocessed dataset to obtain several Bootstrap training subsets; The Bootstrap training subset is input into the IT2FFR main model for calculation to obtain the main model prediction output; The prediction output of the main model is input into the LSTM compensation model for compensation calculation to obtain the corrected prediction result. Based on the corrected prediction results, the LSTM compensation model is iteratively trained using a loss function to obtain the modeling model of the flame combustion line position in the furnace.

5. The method for modeling the flame combustion line position in an urban solid waste incineration furnace according to claim 2, characterized in that, The combustion line calibration results are obtained by using GAN to perform combustion line calibration on the single-frame flame image set, including: The single-frame flame image set is used to extract features using a combustion line edge feature extraction algorithm to obtain a combustion line feature map; the expression of the combustion line edge feature extraction algorithm is as follows: ;in, Image of the edge features of the burning line; This is a morphological processing algorithm; The input is a flame image; The threshold for inverse binarization; Features of the edge of the combustion line; This indicates that 40 expansion operations will be performed. Median filtering; For inverse binarization; The mean value of the combustion line is obtained by calculating the mean value of the vertical coordinate of the white pixels in the combustion line feature map; The variance of the burning line is obtained by calculating the variance of the ordinate of the white pixels in the burning line feature map based on the mean of the burning line. When the variance of the combustion line is less than the variance threshold, the mean of the combustion line is set as the calibration result of the combustion line; when the variance of the combustion line is greater than the variance threshold, the calibration result of the combustion line is set to 0.

6. The method for modeling the flame combustion line position in an urban solid waste incineration process according to claim 4, characterized in that, The Bootstrap training subset is input into the IT2FFR main model for computation to obtain the main model's prediction output, including: Construct an optimization problem; the expression of the optimization problem is: ;in, For the first The first of the trees The optimal splitting threshold for each node; This indicates that the objective function is minimized. function; The number of node samples; , These represent the number of samples in the left subset and the number of samples in the right subset after the split, respectively. Let the mean square error function be used. , These are the true value vectors of the left subset and the right subset, respectively. Construct upper and lower bounds for membership degrees; the expressions for the upper and lower bounds for membership degrees include: , ;in, For interval type II fuzzy sets The upper bound of the membership function; For interval type II fuzzy sets The lower bound membership function; For the first One input variable; The center of the membership function; The width of the membership function; For uncertainty factors; Construct an interval prediction formula; the interval prediction formula includes: , ;in, ; , These represent the lower and upper limits of the output interval for type II fuzzy rule inference, respectively. The number of fuzzy rules for the tree; For the first The normalized activation strength of the rule; , These are the lower and upper limits of the intercept term interval for the consequent of the rule, respectively; , These are the lower and upper limits of the characteristic coefficient interval, respectively; The number of features randomly selected for each tree; Index for features or variables; Based on the Bootstrap training subset, ridge regression is used to iteratively solve the optimization problem, the membership degree upper and lower bounds, and the interval prediction formula to obtain the optimal weights. The main model prediction formula is calculated based on the optimal weights to obtain the main model prediction output; the main model prediction formula is: ;in, ; The predicted output for the main model; The final normalized prediction output of the main model; The standard deviation of the output variable; This is the mean of the output variable; Interval coefficients; The number of trees; The first fusion weight vector Each component.

7. The method for modeling the flame combustion line position in an urban solid waste incineration furnace according to claim 5, characterized in that, The combustion line calibration results are dynamically quantized using SCNN to obtain the quantized values ​​of the flame combustion line position, including: The average value of the combustion line is matched with the template library to obtain a matching template; The maximum similarity is obtained by calculating the similarity between the burning line feature map and the matching template using SCNN and filtering the maximum value. When the maximum similarity is greater than the similarity threshold, the weighted algorithm is used to solve the quantization value of the combustion line in the combustion line feature map to obtain the quantization value of the flame combustion line position. When the maximum similarity is less than the similarity threshold, the average value of the combustion line is set as the quantized value of the flame combustion line position.

Images