Gasifier outlet carbon monoxide dynamic prediction method based on a mixture model

Abstract

This invention discloses a method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model, comprising: Step 1, collecting and preprocessing historical operating data from the gasifier's DCS system to obtain a training set and a test set; Step 2, constructing a parallel weighted fusion model; Step 3, training the parallel weighted fusion model using the training set and determining the optimal hyperparameters; Step 4, inputting the test set into the optimized parallel weighted fusion model to obtain the prediction results; Step 5, inputting real-time collected gasifier operating data into the trained parallel weighted fusion model to output the dynamic predicted value of carbon monoxide at the system outlet. This invention ensures that the model fits historical observation data through a parallel weighted fusion method, conforms to the basic physical laws of carbon element conversion, reduces dependence on a single model, improves the generalization ability of the parallel weighted fusion model under complex operating conditions, and guarantees the rationality and reliability of the prediction results.

Description

Technical Field

[0001] This invention relates to a method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model, belonging to the field of coal chemical product measurement technology. Background Technology

[0002] Coal gasification is a chemical process that converts solid coal into combustible gas. In coal gasification, coal is heated to a high temperature and then undergoes a gasification reaction under oxygen-deficient or rarefied conditions to produce a mixture of gases containing carbon monoxide, hydrogen, methane, and other hydrocarbons. This mixture is commonly referred to as syngas or coal gas.

[0003] The gasifier is one of the core pieces of equipment in the coal gasification process, used to convert solid coal into syngas. Fluidized bed gasification is currently the most prevalent gasification technology, offering advantages such as clean gaseous products and high carbon conversion rates. However, the high-temperature environment inside the gasifier and limitations in measurement methods prevent researchers and producers from accurately obtaining the content of various gases at the gasifier outlet, directly impacting the efficiency of subsequent chemical processes.

[0004] Current prediction methods are divided into data-driven modeling and mechanistic modeling. Data-driven models have the ability to approximate arbitrary functions and do not require detailed knowledge of process mechanisms, exhibiting strong generalization ability. However, data-driven models often suffer from overfitting or underfitting when faced with excessively large operating windows and unexpected situations, potentially leading to non-physical results such as "the total input carbon is much less than the total output carbon," which could pose significant safety risks. The gasification process in a gasifier is extremely complex, and current mechanistic models have all adopted varying degrees of simplification, thus inevitably reducing their accuracy in practical applications. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the defects of the prior art and provide a dynamic prediction method for carbon monoxide at the gasifier outlet based on a hybrid model. The core innovative ideas are twofold: First, a parallel weighted fusion architecture is constructed, in which the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model are operated independently and in parallel. The carbon conservation mechanism model derives the theoretical value of CO based on the carbon conservation law, while the CNN-KAN-Attention data-driven model captures the nonlinear correlation between the input parameters and the measured CO value. The two complement each other. Second, the first theoretical value of CO is obtained using the carbon conservation mechanism model, and the second theoretical value of CO is output using the CNN-KAN-Attention data-driven model. The output of the carbon conservation mechanism model and the output of the CNN-KAN-Attention data-driven model are weighted and fused by a dynamic weight coefficient α to obtain the predicted value of CO, where the predicted value of CO = α × first theoretical value of CO + (1-α) × second theoretical value of CO. This invention employs an architecture of "independent operation of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model + weight fusion," achieving a deep coupling between the data-driven ability to fit complex nonlinear data and the physical rationality of the carbon conservation mechanism model. This allows the model to both closely match historical operating data and strictly adhere to the fundamental laws of carbon transformation. This invention overcomes the limitations of a single model, ensuring the physical interpretability of prediction results while improving generalization ability, accuracy, and stability under complex operating conditions.

[0006] This invention provides a method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model, comprising: Step 1, collecting historical operating data from the gasifier DCS system, the historical operating data including coal characteristic parameters and CO volume fraction at the gasifier outlet at historical times, forming a dataset; preprocessing the dataset, dividing the preprocessed dataset into a training set and a test set; Step 2, constructing a parallel weighted fusion model, the parallel weighted fusion model including a parallel carbon conservation mechanism model and a CNN-KAN-Attention data-driven model; Step 3, using the coal characteristic data at historical times in the training set as the input feature values ​​of the parallel weighted fusion model, and using the CO volume fraction at the gasifier outlet at historical times as the output of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model, constructing a mapping between the input feature values ​​and the CO volume fraction at the gasifier outlet. The optimal model parameters for both the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model are determined by weighting the outputs of the two models using dynamic weight coefficient α, resulting in a trained parallel weighted fusion model. Step 4 involves inputting the test set into the trained parallel weighted fusion model and using it to output prediction results. The prediction results are then denormalized, and an evaluation function is used to assess whether the prediction performance of the parallel weighted fusion model under different operating conditions meets the preset requirements. Step 5 involves inputting the real-time gasifier operation data into the parallel weighted fusion model to output the predicted dynamic carbon monoxide content at the gasifier outlet.

[0007] Priority is given to the gasifier's operating data, which includes coal characteristic parameters, gasifier operating parameters, and quench system monitoring parameters.

[0008] Firstly, the dataset is preprocessed, including: using a 3σ-rule to identify missing outliers in the dataset, and replacing the outliers using the following formula: , , , In the formula, This represents the rate of change of historical operating data between time t and time t-1. For historical running data at time t-1, express The average rate of change before time t, Let t represent the standard deviation at time t, where the sampling times are t=2,…,n, and t' is the initial time.

[0009] Prior to this, in step 1, at least one day of historical operating data is collected, including the complete load change cycle of the gasifier DCS system.

[0010] Prior to this, based on the conservation of carbon mass, the first theoretical value of CO is calculated using the carbon conservation mechanism model. The CNN-KAN-Attention data-driven model adopts the CNN-KAN-Attention architecture and outputs the second theoretical value of CO. The output of the carbon conservation mechanism model and the output of the CNN-KAN-Attention data-driven model are weighted and fused using a weighted fusion equation with dynamic weight coefficient α to obtain the predicted value of CO. The weighted fusion equation is: predicted value of CO = α × first theoretical value of CO + (1-α) × second theoretical value of CO.

[0011] Prior to this, the CNN-KAN-Attention data-driven model consists of an input layer, a CNN layer, a KAN layer, an attention mechanism layer, and an output layer connected in sequence.

[0012] Prior to this, in step 3, the hyperparameters of the parallel weighted fusion model are set, including: batch size of 128, number of iterations of 200, optimizer of Adam, learning rate of 0.005, and learning rate decay factor of 0.5.

[0013] Prior to this, a dynamic prediction system for carbon monoxide at the gasifier outlet based on a mechanism and data-driven model includes: a data acquisition module for acquiring historical operating data from the gasifier DCS system, the historical operating data including coal characteristic parameters, gasifier operating parameters, and quench system detection parameters at historical times, forming a dataset; a preprocessing module for preprocessing the dataset, dividing the preprocessed dataset into a training set and a test set; a fusion model construction module for constructing a parallel weighted fusion model that combines an element conservation mechanism model and a CNN-KAN-Attention data-driven model; and a dynamic weight fusion module for using the coal characteristic data at historical times included in the training set as input feature values ​​to the parallel weighted fusion model, and using the CO volume fraction at the gasifier outlet at historical times as the output of the carbon element conservation mechanism model and the CNN-KAN-Attention data-driven model, constructing a mapping relationship between the input feature values ​​and the CO volume fraction at the gasifier outlet, and utilizing dynamic weight fusion. The outputs of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model are weighted and fused using a weighted coefficient α to determine the optimal model parameters for both models, resulting in a trained parallel weighted fusion model. The weighted fusion equation is: CO prediction value = α × first theoretical CO value + (1-α) × second theoretical CO value. The test set is input into the trained parallel weighted fusion model, which then outputs the prediction results. The prediction results are inversely normalized, and an evaluation function is used to judge whether the prediction performance of the parallel weighted fusion model under different operating conditions meets the preset requirements. The output module is optimized so that if the parallel weighted fusion model meets the preset requirements, the real-time gasifier operation data is input into the parallel weighted fusion model that meets the preset requirements, and the model outputs the dynamic predicted value of carbon monoxide content at the gasifier outlet.

[0014] Preferably, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described herein.

[0015] Preferably, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described herein.

[0016] The beneficial effects achieved by this invention are as follows:

[0017] This invention proposes a dynamic prediction method for carbon monoxide at the gasifier outlet based on a hybrid model. It establishes a model that can accurately predict the gasifier outlet products, providing theoretical value and practical guidance for achieving stable and efficient operation of the production system. Through the advantages of parallel fusion mechanism and data-driven approach, it achieves accurate prediction of CO concentration. This invention achieves two innovations: First, it constructs a parallel weighted fusion architecture, independently and in parallel operating the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model. The carbon conservation mechanism model derives the theoretical value of CO based on the carbon conservation law, while the CNN-KAN-Attention data-driven model captures the nonlinear correlation between input parameters and CO measurement values, with the two complementing each other's advantages. Second, it uses the carbon conservation mechanism model to obtain the first theoretical value of CO, and the CNN-KAN-Attention data-driven model, using the CNN-KAN-Attention architecture, outputs the second theoretical value of CO. By dynamically weighting the output of the carbon conservation mechanism model and the output of the CNN-KAN-Attention data-driven model with an α weighted fusion, the predicted value of CO is obtained, where the predicted value of CO = α × first theoretical value of CO + (1-α) × second theoretical value of CO. This invention employs an architecture of "independent operation of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model + weight fusion," achieving a deep coupling between the data-driven model's ability to fit complex nonlinearities and the physical rationality of the carbon conservation mechanism model. This ensures the model closely matches historical operational data while strictly adhering to the fundamental laws of carbon transformation. This invention overcomes the limitations of single models, guaranteeing the physical interpretability of prediction results while improving generalization ability, accuracy, and stability under complex operating conditions. This invention designs a method and system for dynamic prediction of carbon monoxide at gasifier outlets, capable of adapting to multivariate and nonlinear characteristics under complex operating conditions, exhibiting strong robustness, interpretability, and generalization ability, with high computational efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2 This is a typical convolutional structure diagram.

[0021] Figure 3 This is a structural diagram of the KAN layer.

[0022] Figure 4This is a schematic diagram of the attention mechanism.

[0023] Figure 5 This is a graph showing the prediction results of the pure data model CNN-KAN-Attention.

[0024] Figure 6 This is a graph showing the prediction results of the parallel weighted fusion model. Detailed Implementation

[0025] See Figure 1 This embodiment provides a method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model, including the following steps: Step 1: The gasifier adopts a two-stage dry pulverized coal gasifier, and data is collected and preprocessed; Step 1.1: Collect gasifier operation data from the gasifier DCS system: According to the gasifier mechanism, the collected gasifier operation data includes coal characteristic parameters, gasifier operation parameters, and quench system detection parameters, specifically including 13 sets of characteristic values ​​such as carbon element mass fraction, pulverized coal feed rate, steam flow rate, total syngas flow rate, and methane volume fraction, as detailed in Table 1.

[0026] The data for this case study comes from the DCS system of a gasifier unit in a chemical plant in Tianjin. To verify the model's performance in short-term dynamic prediction, typical 24-hour data from 00:00 on June 13, 2025 to 00:00 on June 14, 2025 was selected as the dataset. Data was collected every 10 seconds, resulting in 8640 sample data points. This dataset includes various operating conditions such as load increases, load decreases, and drastic load fluctuations, encompassing a complete load change cycle of a gasifier DCS system, and is highly representative.

[0027] Table 1. Eigenvalues

[0028]

[0029] Step 1.2: Preprocess the obtained gasifier operation data: Due to the complex environment of the gasifier production site and the presence of noise in the signal, the original gasifier operation data stored in the gasifier DCS system contains anomalies of varying degrees. This embodiment uses a 3σ-rule to detect anomalies in the original gasifier operation data and replaces the anomalies based on the following formula: , , , ,in, Representing auxiliary variables The rate of change at adjacent times t and t-1, express The average rate of change before time t, The standard deviation at time t is denoted as t=2,…,n, where t=2,…,n are the sampling times.

[0030] The gasifier operating data, after replacing the abnormal data, was linearly normalized using the following formula to standardize the data dimensions: In the formula: X n Let X be the normalized data, and X be the current data to be normalized. min X is the minimum value of the total data to be normalized. max This represents the maximum value of the total data to be normalized.

[0031] Step 1.3: Dataset Splitting: The original 8640 sample data points are split into 8640 standard input matrices. Each matrix corresponds to a label ordered from 1 to 8640, ensuring that the standard input matrix and the matrix index are consistent, i.e., the indexes of the standard input matrix and the labels all start from 1. All the split sample data are then divided into training and test sets in an 8:2 ratio, maintaining the original sampling order.

[0032] Step 2: Construct a parallel weighted fusion model, adopting an architecture of "independent operation of the mechanism model and the data-driven model + weight fusion". Sub-model 1 is a CNN-KAN-Attention data-driven model, and sub-model 2 is a carbon conservation mechanism model. The outputs of the two models are fused by weighting the dynamic weight coefficient α. Step 2.1: The CNN-KAN-Attention data-driven model includes an input layer, a CNN layer, a KAN layer, an Attention layer, and an output layer connected in sequence. The CNN layer extracts local features, and the KAN layer processes high-order nonlinear relationships. The combination of the two captures the complex characteristics of CO generation. The Attention layer filters features and strengthens the influence of key parameters (such as oxygen-coal ratio and temperature) on CO prediction, improving model accuracy and adaptability to operating conditions. This architecture, through progressive processing of "local feature extraction, nonlinear mapping, and key information focusing", deeply adapts to the strong nonlinear dynamic characteristics of CO generation in gasifiers.

[0033] CNN stands for Convolutional Neural Network. A Convolutional Neural Network is a type of neural network specifically designed to process data with a grid-like structure. It includes an input layer, convolutional layers, activation layers, pooling layers, and fully connected layers, such as... Figure 2 As shown in (c).

[0034] Convolutional layers are the core of CNN layers. Each convolutional layer consists of multiple convolutional kernels, each of which acts like a filter, primarily extracting features from the input data. Through convolution operations, convolutional layers can effectively capture important features in the data and enhance key features in the original signal, while effectively reducing the dimensionality of the features. Figure 2Figure (a) shows the process of a 4×4 input being processed by a 3×3 convolution kernel and the output result. The convolution kernel moves one unit to the right each time, simultaneously performing an inner product with the data window. Taking the first window as an example, the calculation process is as follows: In a CNN layer, after each convolution of a portion of the data matrix, the data matrix continuously moves in parallel until all the data has been convolved. Figure 2 As can be seen in (a), through the action of the 3×3 convolution kernel, the original two-dimensional data is transformed from the initial 4×4 into a 2×2 two-dimensional matrix.

[0035] Pooling layers are typically connected after convolutional layers. Their main function is to perform secondary data extraction, reducing the amount of data needed for prediction and thus lowering the network's complexity. Pooling operations mainly include average pooling and max pooling. Average pooling extracts the average value within a given region, while max pooling extracts the maximum value within a given region. Figure 2 As shown in (b).

[0036] Activation layers enhance the nonlinearity of the model by applying activation functions to the outputs of fully connected layers or convolutional layers. This invention chooses the ReLU function: It exhibits stronger robustness and has advantages such as high speed, good sparsity, and good derivative properties.

[0037] Fully connected layers are located at the end of CNN layers, connecting with neurons in the upper layers to transform the data dimensions while preserving useful information. Their main function is to learn the output features of previous layers, perform global integration, and map them to the sample label space.

[0038] The KAN layer comprises a KAN neural network, constructed based on the Kolmogorov-Arnold representation theorem. Figure 3 This is a schematic diagram of the KAN neural network structure. Figure 3 As can be seen, the input layer receives the raw data, and the internal functions on the connection lines between it and the hidden layer... ( Figure 3 A small rectangle with a curved pattern is used as a learnable activation function, which adaptively adjusts during training. This function is applied to the transformed input data and then summed. Operations are passed to the hidden layer; the hidden layer further processes the data to extract features, and the external function on the connection line between the hidden layer and the output layer... (Small rectangles with curved patterns) After further data transformation, the data is passed to the output layer, which then outputs the final result. Unlike traditional multilayer perceptrons (MLPs) that place activation functions at the nodes, KAN neural networks place learnable activation functions at the network edges (weight positions). They achieve high expressive power and parameter efficiency through spline function fitting. KAN layers are used to deeply mine complex nonlinear relationships in input features. KAN layers Including external functions and internal functions, represented as: Where x is an n-dimensional input vector, representing the input to... The data contains n components x1, x2, ..., xn. n Where x p This represents the p-th component in the input vector x, where p ranges from 1 to n and is used to index the elements of the input vector. For a learnable intrinsic function, there are two subscripts, q and p. p is used to index the components of the input vector, and q is the number of the different intrinsic functions. The domain is [0,1], and the range is the set of real numbers R; It is a learnable external function, with its domain and range both being the set of real numbers R; f represents the expression with respect to x1, x2, ..., xn. n A function with variables.

[0039] The attention layer employs an attention mechanism, which borrows from how humans concentrate and process information, much like the brain's information filtering process. Deep learning based on this attention mechanism aims to teach machines to distinguish between important and unimportant parts of data. Its structure is as follows: Figure 4 As shown, the calculation steps are as follows: 1) For the input sequence [y1,y2,...,y...], ... n Encode the data to obtain a set of query vectors q; then, use the similarity between the scores [s1, s2, ..., s...] to obtain a set of query vectors q. n The scoring function is as follows: In the formula, V, W, and U are matrices of parameters to be learned. 2) Normalization is performed using the Softmax function to obtain the input vector a for each group. i The formula is as follows: 3) Calculate the weighted sum based on the weights and their corresponding value vectors to obtain the final hidden layer state value output a: .

[0040] Step 2.2: The two-stage dry pulverized coal gasifier achieves efficient coal conversion through "segmented reaction": the first stage is a high-temperature oxidation zone, where pulverized coal reacts violently with oxygen to release heat and generate gases such as CO2 and CO, which provide energy for the system; the second stage is a medium-temperature reduction zone, where pulverized coal reacts with CO2 and water vapor generated in the first stage to further generate CO, ultimately increasing the CO content in the syngas. The entire process follows the law of conservation of carbon in oxidation-reduction reactions.

[0041] (1) Oxidation reaction: Carbon and oxygen burn completely to produce carbon dioxide. When the amount of oxygen is insufficient, carbon undergoes incomplete combustion to produce carbon monoxide. , .

[0042] (2) Reduction reaction: In the medium-low temperature region, carbon dioxide reacts with carbon at high temperature to produce carbon monoxide; water vapor reacts with carbon to produce water gas, producing carbon monoxide and hydrogen. In addition, side reactions such as methanation also occur. , , .

[0043] (3) Based on the above gasifier operation mechanism and chemical reaction, and ignoring residual carbon, a carbon conservation mechanism model of "total input = total output" is constructed by quantitatively analyzing the input and output of carbon elements, as shown in the following formula: , In the formula: The feed rate of pulverized coal is (kg / h). The feed rate of pulverized coal in the second stage is (kg / h). The percentage of carbon in pulverized coal (%) This represents the total syngas flow rate (Nm³ / h). The CO volume fraction (%) is predicted by the carbon conservation mechanism model. CO2 volume fraction (%) The volume fraction of methane is (%). The density of CO under standard conditions. The density of CO2 under standard conditions. This represents the density of methane under standard conditions.

[0044] According to the law of conservation of mass, we can simultaneously solve... After sorting, we get: .

[0045] Step 2.3, Dynamic Weight Fusion Layer: The outputs of the two models are fused using weight coefficients α (0 ≤ α ≤ 1). Then the total loss function is: In the formula, L represents the total loss. The CO volume fraction for the i-th data point predicted by the parallel weighted fusion model. The CO volume fraction predicted by the CNN-KAN-Attention data-driven model for the i-th data point. Let α be the actual measured CO volume fraction of the i-th data point, α be the weighting coefficient, and N be the number of data points.

[0046] α optimization strategy: Dynamically adjust the dynamic weight coefficient α to find the minimum value of L. When the coal quality fluctuation is >5%, the dynamic weight coefficient α increases to 0.6~0.8, relying on the mechanism to ensure physical rationality; when the coal quality is stable, the dynamic weight coefficient α decreases to 0.2~0.4, relying on the data model to improve accuracy.

[0047] Step 3: Train the parallel weighted fusion model using the training set. The parallel weighted fusion model incorporates the Dropout strategy to suppress overfitting and adopts the Adam optimization algorithm to adjust the model parameters. Gradient descent optimization is performed on the model parameters and mechanistic parameters to update the parameters and minimize the total loss. Repeat the above steps, adjusting the parameters until the model converges to a suitable state. This model is implemented using MATLAB R2023b, on a computer equipped with an AMD Ryzen 5 3500U processor with a clock speed of 1.6GHz, 12.0GB of RAM, a Radeon Vega graphics card, and 16GB of RAM. The LearnRateDropFactor is 0.5, the initial learning rate is 0.005, the batch size is 200, the LearnRateDropPeriod is 150, and α is 0.3. The evaluation functions include Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). 2 .

[0048] .

[0049] In the formula, y i , ȳi represent the actual measured value, predicted value, and mean of the actual measured value of the test set, respectively, and n represents the total number of samples. Meeting the preset requirements includes: 1. Mean Absolute Percentage Error (MAPE) within the preset MAPE threshold range; 2. Root Mean Square Error (RMSE) within the preset RMSE threshold range; 3. Mean Absolute Error (MAE) within the preset MAPE threshold range; 4. Coefficient of Determination within the preset coefficient of determination range.

[0050] Step 4: Input the test set into the optimized parallel weighted fusion model, obtain the prediction results, and analyze and evaluate the prediction results using the evaluation function to determine whether the model's prediction performance meets the requirements under different operating conditions. To evaluate the superiority of the hybrid model in predicting carbon monoxide at the gasifier outlet, simulation experiments were conducted using the same dataset, employing both the pure data-driven CNN-KAN-Attention model and the hybrid model. The comparison results of each model are shown below. Figure 5 and Figure 6 The evaluation index values ​​for each model are shown in Table 2.

[0051] Table 2. Comparison of prediction results from different models

[0052]

[0053] from Figure 5 and Figure 6 It can be seen that in the carbon monoxide prediction model at the gasifier outlet, when the target value is relatively stable, the predictions of various models can closely follow the actual value with small fluctuations. However, the parallel weighted fusion model performs better than the pure data-driven CNN-KAN-Attention model. When the carbon monoxide value changes significantly, the predictions of the parallel weighted fusion model closely follow the actual value and quickly stabilize after the change, while the predictions of other models sometimes lag behind the actual value and fluctuate greatly. Table 2 shows that the parallel weighted fusion model performs better than the pure data-driven CNN-KAN-Attention model. The MAE, RMSE, and MAPE of the parallel weighted fusion model are 0.029474, 0.040803, and 0.06901%, respectively, which are 27.91%, 26.17%, and 20.43% lower than those of the pure data-driven CNN-KAN-Attention model. The R-value of the parallel weighted fusion model... 2 The accuracy was 98.856%, a 2.62% improvement compared to the pure data-driven CNN-KAN-Attention model. This indicates that the parallel weighted fusion model combines the advantages of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model, further improving the accuracy, stability, and interpretability of the parallel weighted fusion model. Step 5: Input the real-time collected gasifier operation data into the trained parallel weighted fusion model to output the predicted dynamic carbon monoxide value at the gasifier outlet.

[0054] In this embodiment of the application, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0055] In this application embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0056] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0057] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention described herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not invented herein. The specification and embodiments are to be considered exemplary only.

[0058] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

Claims

1. A method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model, characterized in that, include: Step 1: Collect historical operating data from the gasifier DCS system. The historical operating data includes the characteristic parameters of coal at historical times and the CO volume fraction at the gasifier outlet at historical times, forming a dataset. Preprocess the dataset and divide the preprocessed dataset into a training set and a test set. Step 2: Construct a parallel weighted fusion model, which includes a parallel carbon conservation mechanism model and a CNN-KAN-Attention data-driven model; the expression for the carbon conservation mechanism model is: In the formula, This refers to the coal powder feed rate. This refers to the two-stage pulverized coal feed rate. This represents the mass fraction of carbon in pulverized coal. This represents the total syngas flow rate. The CO volume fraction predicted by the carbon conservation mechanism model. This represents the volume fraction of CO2. This represents the volume fraction of methane. The density of CO under standard conditions. The density of CO2 under standard conditions. The density of methane under standard conditions; Based on the conservation of carbon mass, the first theoretical value of CO is calculated using the carbon conservation mechanism model. The CNN-KAN-Attention data-driven model adopts the CNN-KAN-Attention architecture and outputs the second theoretical value of CO. The output of the carbon conservation mechanism model and the output of the CNN-KAN-Attention data-driven model are weighted and fused using a weight fusion equation with dynamic weight coefficient α to obtain the predicted value of CO. The weight fusion equation is: predicted value of CO = α × first theoretical value of CO + (1-α) × second theoretical value of CO. Step 3: Use the coal characteristic data from historical moments in the training set as the input feature values ​​of the parallel weighted fusion model, and use the CO volume fraction at the gasifier outlet at historical moments as the output of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model. Construct a mapping relationship between the input feature values ​​and the CO volume fraction at the gasifier outlet. Use the dynamic weight coefficient α to weight and fuse the outputs of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model to determine the optimal model parameters of the carbon conservation mechanism model and the optimal model parameters of the CNN-KAN-Attention data-driven model, and obtain the parallel weighted fusion model after training. Step 4: Input the test set into the trained parallel weighted fusion model, and use the parallel weighted fusion model to output the prediction results; perform inverse normalization on the prediction results, and use the evaluation function to judge the prediction results to determine whether the prediction performance of the parallel weighted fusion model under different working conditions meets the preset requirements. Step 5: If the parallel weighted fusion model meets the preset requirements, the real-time gasifier operation data is input into the parallel weighted fusion model that meets the preset requirements, and the dynamic content prediction value of carbon monoxide at the gasifier outlet is output using the parallel weighted fusion model.

2. The method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model according to claim 1, characterized in that, The gasifier operation data includes coal characteristic parameters, gasifier operation parameters, and quench system monitoring parameters.

3. The method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model according to claim 1, characterized in that, Preprocessing of the dataset includes: The missing outliers in the dataset are identified using the 3σ-rule, and the outliers are replaced using the following formula: ， ， ， ， In the formula, This represents the rate of change of historical operating data between time t and time t-1. For historical running data at time t-1, express The average rate of change before time t, Let t represent the standard deviation at time t, where the sampling times are t=2,…,n, and t' is the initial time.

4. The method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model according to claim 1, characterized in that, In step 1, at least one day of historical operating data is collected, including the complete load change cycle of the gasifier DCS system.

5. The method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model according to claim 1, characterized in that, The CNN-KAN-Attention data-driven model consists of an input layer, a CNN layer, a KAN layer, an attention mechanism layer, and an output layer connected in sequence.

6. The method for dynamic prediction of carbon monoxide at the gasifier outlet based on a hybrid model according to claim 1, characterized in that, In step 3, the hyperparameters of the parallel weighted fusion model are set, including: batch size of 128, number of iterations of 200, optimizer of Adam, learning rate of 0.005, and learning rate decay factor of 0.

5.

7. A dynamic prediction system for carbon monoxide at gasifier outlet based on mechanism and data-driven model, employing the dynamic prediction method for carbon monoxide at gasifier outlet based on a hybrid model as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to collect historical operating data from the gasifier DCS system. The historical operating data includes coal characteristic parameters, gasifier operating parameters, and quench system detection parameters at historical moments, forming a dataset. The preprocessing module is used to preprocess the dataset, dividing the preprocessed dataset into training and testing sets; The fusion model building module is used to construct a parallel weighted fusion model that runs in parallel with the element conservation mechanism model and the CNN-KAN-Attention data-driven model; The dynamic weighted fusion module uses historical coal characteristic data from the training set as input feature values ​​to the parallel weighted fusion model, and historical CO volume fraction at the gasifier outlet as the output of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model. It constructs a mapping relationship between the input feature values ​​and the CO volume fraction at the gasifier outlet, and uses dynamic weight coefficient α to weight and fuse the outputs of the carbon conservation mechanism model and the CNN-KAN-Attention data-driven model to determine the optimal model parameters for both models, thus obtaining the trained parallel weighted fusion model. The weighted fusion equation is: CO predicted value = α × first theoretical CO value + (1-α) × second theoretical CO value. The test set is input into the trained parallel weighted fusion model, and the parallel weighted fusion model is used to output the prediction results. The prediction results are denormalized, and the evaluation function is used to judge the prediction results to determine whether the prediction performance of the parallel weighted fusion model under different working conditions meets the preset requirements. The optimized output module is used to input the real-time gasifier operation data into the parallel weighted fusion model that meets the preset requirements if the parallel weighted fusion model meets the preset requirements, and then output the predicted dynamic content of carbon monoxide at the gasifier outlet using the parallel weighted fusion model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 6.

Images