A method and system for dynamic prediction of hydrogen at the outlet of a two-stage gasifier

Abstract

This invention discloses a method and system for dynamic prediction of hydrogen at the outlet of a two-stage gasifier, comprising the following steps: collecting historical operating data of the gasifier; constructing a derived feature dataset based on the historical operating data; preprocessing the derived feature dataset and dividing it into a training set and a test set; constructing a deep learning model, which sequentially includes a convolutional neural network layer, a bidirectional gated recurrent unit layer, and an attention mechanism layer; embedding physical constraint terms into the loss function of the deep learning model to construct a total loss function; training the deep learning model using the training set to obtain an optimized model; inputting the test set into the optimized model and analyzing and evaluating it through an evaluation function; obtaining real-time operating data and inputting it into the optimized model to output the real-time predicted value of hydrogen at the gasifier outlet; thereby solving the problem that existing models cannot meet the requirements of high accuracy and physical interpretability.

Description

Technical Field

[0001] This invention relates to the field of coal chemical process parameter measurement technology, and in particular to a method and system for dynamic prediction of hydrogen at the outlet of a two-stage gasifier. 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 most crucial 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. Therefore, it is necessary to establish a model that can accurately predict the gasifier outlet products (taking hydrogen as an example), providing theoretical value and practical guidance for achieving stable and efficient operation of the production system.

[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 unforeseen circumstances, potentially leading to non-physical results such as "the total input hydrogen is much less than the total output hydrogen," 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.

[0005] Therefore, there is an urgent need in this field for a hybrid modeling method that can combine the physical rigor of mechanistic models with the adaptive advantages of data-driven models, in order to achieve high-precision, physically reasonable, and real-time reliable soft measurement of hydrogen content at the gasifier outlet. Summary of the Invention

[0006] The technical problem solved by this invention is that pure data-driven models have poor physical rationality and insufficient generalization ability when predicting hydrogen content at the gasifier outlet, while pure mechanistic models have low prediction accuracy due to simplified assumptions. Both are difficult to meet the common requirements of industrial applications for high accuracy, strong robustness and physical interpretability.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for dynamic prediction of hydrogen at the outlet of a two-stage gasifier, comprising the following steps:

[0008] Step S1: Collect historical operating data of the gasifier, construct at least one derived feature dataset based on the historical operating data, preprocess the derived feature dataset, and divide it into a training set and a test set; the historical operating data includes coal quality characteristic parameters and gasifier operating parameters.

[0009] Step S2: Construct a deep learning model, which sequentially includes a convolutional neural network layer, a bidirectional gated recurrent unit layer, and an attention mechanism layer; embed a physical constraint term based on the conservation of hydrogen mass into the loss function of the deep learning model to construct the total loss function.

[0010] Step S3: Train the deep learning model using the training set to obtain an optimized model.

[0011] Step S4: Input the test set into the optimization model to obtain the predicted verification value, and analyze and evaluate the predicted verification value through the constructed evaluation function.

[0012] Step S5: Obtain real-time operating data of the gasifier, input it into the optimization model, and output the real-time predicted value of hydrogen at the gasifier outlet.

[0013] Preferably, the derived features include at least one of water-carbon ratio, oxygen-coal ratio, and oxygen / steam pressure ratio; the water-carbon ratio is calculated from the total steam volume of the first-stage burner, the first-stage coal powder feed rate, and the carbon element mass fraction in the coal powder; the oxygen-coal ratio is calculated from the total oxygen flow rate of the first-stage burner and the first-stage coal powder feed rate; and the oxygen / steam pressure ratio is calculated from the oxygen pressure and the steam pressure.

[0014] Preferably, the formula for calculating the water-to-carbon ratio is: ;in, Indicates the water-to-carbon ratio. This indicates the flow rate of steam entering a section of the burner. This represents the pulverized coal feed mass flow rate of a gasification chamber. This indicates the mass fraction of carbon in pulverized coal.

[0015] The formula for calculating the oxygen-to-coal ratio is: ;in, Indicates the oxygen-to-coal ratio. This indicates the oxygen volumetric flow rate of a section of the vaporization chamber.

[0016] The formula for calculating the oxygen / vapor pressure ratio is as follows: ;in, The oxygen / vapor pressure ratio, This indicates the pressure at which oxygen enters the burner. This indicates the pressure at which steam enters the burner.

[0017] Preferably, in step S1, the preprocessing specifically includes missing data processing, data dimensionality reduction, and data normalization; wherein, the missing data processing uses 3σ-rule to detect abnormal data and replaces it based on the rate of change of data at adjacent time points.

[0018] Preferably, in step S2: the physical constraint term based on the conservation of hydrogen mass is calculated from the deviation between the total input and output of hydrogen; the total loss function is the weighted sum of the data fitting loss term and the physical constraint term.

[0019] Preferably, the total hydrogen input is calculated based on the coal powder feed rate, the hydrogen and moisture content in the coal, and the steam flow rate; the total hydrogen output is calculated based on the total syngas flow rate, the predicted hydrogen gas integral number, and the methane volume fraction.

[0020] Preferably, in step S1, the collected historical operating data includes coal quality characteristic parameters, gasifier operating parameters, and some detection parameters of the quench system.

[0021] Preferably, the convolutional neural network layer is used to extract local correlation features of the input parameters; the bidirectional gated recurrent unit layer is used to capture the dynamic temporal pattern of hydrogen content; and the attention mechanism layer is used to strengthen the influence weight of key parameters on the prediction results.

[0022] A dynamic prediction system for hydrogen at the outlet of a two-stage gasifier includes: a data acquisition and preprocessing module for acquiring real-time gasifier operating data and completing derived feature construction, missing data processing, and data normalization; a feature extraction and temporal modeling module, including a convolutional neural network unit, a bidirectional gated recurrent unit, and an attention mechanism unit connected in sequence, for extracting features from the preprocessed data and modeling; a physical constraint calculation module for calculating the hydrogen mass conservation deviation based on the input data; and a model optimization and output module for optimizing model parameters based on a total loss function including data fitting loss and physical constraint loss, and outputting the final predicted hydrogen content value.

[0023] Preferably, the data acquisition and preprocessing module is specifically used to calculate at least one derived feature among the water-carbon ratio, oxygen-coal ratio, and oxygen / steam pressure ratio.

[0024] The beneficial effects of this invention are as follows: First, by introducing derived features based on the gasification mechanism (such as water-to-carbon ratio and oxygen-to-coal ratio), this invention directly uses the key proportional relationships reflecting the essence of the reaction as model input, effectively solving the problem of collinearity among the original parameters, and enabling the model to more directly capture the core factors affecting hydrogen generation. Simultaneously, the model's hybrid network structure can deeply mine data features from three levels: local, global, and key, comprehensively capturing the nonlinear and dynamic temporal characteristics of the gasification process.

[0025] Secondly, by embedding the conservation of hydrogen mass as a constraint term into the loss function, this invention fundamentally eliminates non-physical phenomena such as "input hydrogen being less than output hydrogen." This not only enhances the physical rationality and credibility of the prediction results, but also plays a crucial guiding and corrective role in scenarios with sparse data, high noise, or operating conditions exceeding the range of training data, preventing the prediction direction from deviating too far and greatly improving the reliability and safety of the model.

[0026] Third, by introducing derived features and physical constraints with clear physical meaning, this invention makes it easier for engineers to understand which key process parameters (such as the water-to-carbon ratio) and physical principles (such as element conservation) the model is based on for prediction. This facilitates mechanism-level analysis and fault diagnosis when prediction deviations occur, and promotes the deep application of artificial intelligence technology in the industrial field. Attached Figure Description

[0027] Figure 1 The flowchart of the dynamic prediction method for hydrogen at the outlet of a two-stage gasifier provided by the present invention is shown.

[0028] Figure 2 The convolutional structure diagram is provided by the present invention for a method for dynamic prediction of hydrogen at the outlet of a two-stage gasifier.

[0029] Figure 3 The diagram shows the GRU structure of a two-stage gasifier outlet hydrogen dynamic prediction method provided by this invention.

[0030] Figure 4 The diagram shows the BiGRU structure of a two-stage gasifier outlet hydrogen dynamic prediction method provided by this invention.

[0031] Figure 5 This invention provides an attention mechanism diagram for a method of dynamically predicting hydrogen at the outlet of a two-stage gasifier.

[0032] Figure 6 This is a sequence diagram of test set samples before the input derived feature values ​​for a two-stage gasifier outlet hydrogen dynamic prediction method provided in an embodiment of the present invention.

[0033] Figure 7 This is a sequence diagram of test set samples after inputting derived feature values ​​for a method for dynamic prediction of hydrogen at the outlet of a two-stage gasifier, as provided in an embodiment of the present invention.

[0034] Figure 8 This is a sequence of test set samples for a CNN-BiGRU-Attention method for dynamically predicting hydrogen output from a two-stage gasifier, as provided in an embodiment of the present invention.

[0035] Figure 9 This is a test set sample sequence diagram of a hybrid model for a method for dynamic prediction of hydrogen at the outlet of a two-stage gasifier, provided in an embodiment of the present invention. Detailed Implementation

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0037] This embodiment uses a two-stage dry pulverized coal gasifier in a chemical plant as an example to implement the soft measurement method for hydrogen content at the gasifier outlet described in this invention. The actual measured value of the hydrogen integral at the gasifier outlet is obtained through periodic sampling and offline chromatographic analysis, which has a lag of several hours and cannot be used for real-time control. The goal of this invention is to establish a measurement model that uses easily measurable DCS parameters to predict the outlet hydrogen content in real time.

[0038] Therefore, the present invention provides a method for dynamic prediction of hydrogen at the outlet of a two-stage gasifier, comprising the following steps: Step S1, data preparation and preprocessing: collecting historical operating data of the gasifier, constructing at least one derived feature dataset based on the historical operating data, preprocessing the derived feature dataset, and dividing it into a training set and a test set; the historical operating data includes coal quality characteristic parameters and gasifier operating parameters.

[0039] Specifically, it includes the following sub-steps: Step S11, data acquisition and derivation feature construction: Based on the gasifier mechanism, the historical operating data collected includes coal quality characteristic parameters, gasifier operating parameters, and some detection parameters of the quench system, including coal element analysis such as hydrogen mass fraction and carbon mass fraction, coal powder feed rate and steam flow rate, syngas total flow rate and methane volume fraction of the quench system, etc. In addition, 3 derivation feature values ​​were constructed, totaling 15 sets, as detailed in Table 1.

[0040] Specifically, historical operating data for 24 consecutive hours from 00:00 on May 23, 2025 to 00:00 on May 24, 2025 were collected from the DCS system of the gasifier, with a sampling interval of 10 seconds, resulting in a total of 8,640 samples. The collected data variables are shown in Table 1, covering three main categories: coal quality characteristics, operating parameters, and quench system parameters, totaling 13 raw variables.

[0041] Specifically, to address the collinearity issue among the original parameters and to more directly reflect the intrinsic mechanism of the gasification reaction, this invention constructs three key derived features based on the original variables in Table 1. These derived features include at least one of the following: water-to-carbon ratio, oxygen-to-coal ratio, and oxygen / steam pressure ratio. Then, these three derived features are merged with the original 13 features to form a final dataset containing a 16-dimensional feature vector.

[0042] The water-to-carbon ratio is calculated from the total steam volume of the first-stage burner, the feed rate of the first-stage pulverized coal, and the mass fraction of carbon in the pulverized coal.

[0043] The oxygen-to-coal ratio is calculated from the total oxygen flow rate of one stage and the coal powder feed rate of the other stage.

[0044] The oxygen / vapor pressure ratio is calculated from the oxygen pressure and the vapor pressure.

[0045] Preferably, the formula for calculating the water-to-carbon ratio is: ;in, Indicates the water-to-carbon ratio. This indicates the steam flow rate (kg / h, measured under standard conditions) entering a section of the burner. This represents the pulverized coal feed mass flow rate (kg / h) of a gasification chamber. This indicates the mass fraction (%) of carbon in pulverized coal, which can be obtained through elemental analysis of coal.

[0046] The formula for calculating the oxygen-to-coal ratio is: ;in, Indicates the oxygen-to-coal ratio. This represents the oxygen volumetric flow rate (Nm³ / h, measured under standard conditions) of a gasification chamber.

[0047] The formula for calculating the oxygen / vapor pressure ratio is as follows: ;in, The oxygen / vapor pressure ratio, This indicates the pressure (MPa) at which oxygen enters the burner. This indicates the pressure (MPa) at which steam enters the burner.

[0048] Table 1: List of collected feature variables .

[0049] Step S12, data preprocessing: Due to the complex industrial environment, there may be outliers in the raw DCS data.

[0050] The preprocessing described in this invention specifically includes missing data processing, data dimensionality reduction, and data normalization; wherein, the missing data processing uses a 3σ-rule to detect abnormal data and replaces it based on the rate of change of data at adjacent time points.

[0051] The specific steps are as follows: ; ; ; ;in, Representing auxiliary variables The rate of change between two adjacent time points express The average rate of change before time t, Let represent the standard deviation; where t=2,…,n are the sampling times.

[0052] After the abnormal data is processed, all feature variables are linearly normalized to compress their scale to the [0,1] interval in order to accelerate the convergence of model training.

[0053] The normalization formula is: Where Xn is the normalized data, X is the current data to be normalized, Xmin is the minimum value of the total data to be normalized, and Xmax is the maximum value of the total data to be normalized.

[0054] Step S13: Dataset partitioning. The preprocessed 8640 sample data (each sample containing 16 features) and their corresponding outlet hydrogen content labels (lagging the corresponding actual analysis values) are kept in chronological order and divided into training and test sets in an 8:2 ratio. The training set contains 6912 samples for model training and parameter tuning; the test set contains 1728 samples for independently evaluating the model's generalization performance.

[0055] Step S2, Physics-Guided Hybrid Model Construction. A deep learning model is constructed, and a physical constraint term based on the conservation of hydrogen mass is embedded into the loss function of the deep learning model to construct the total loss function.

[0056] Step S21, construct the convolutional neural network layer (CNN) (i.e. Figure 2 The structure of the convolutional neural network (CNN) shown in (c) is shown in the figure.

[0057] A convolutional neural network (CNN) is a type of neural network specifically designed to process data with a grid-like structure. It consists of an input layer, convolutional layers, activation layers, pooling layers, and fully connected layers, such as... Figure 2 As shown.

[0058] Convolutional layers are the core of CNNs. Each convolutional layer consists of multiple convolutional kernels, each acting like a filter. Their primary function is to extract 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 convolution operation process of a 4×4 input passing through a 3×3 convolution kernel and outputting the 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: .

[0059] In CNNs, 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), the original two-dimensional data is transformed from a 4×4 matrix into a 2×2 two-dimensional matrix through the action of a 3×3 convolution kernel.

[0060] 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) of the diagram.

[0061] Activation layers enhance the nonlinearity of a model by applying activation functions to the outputs of fully connected layers or convolutional layers. In this embodiment, the present invention selects the ReLU function, which exhibits stronger robustness and has advantages such as high speed, good sparsity, and good derivative properties.

[0062] .

[0063] Fully connected layers are located at the end of a CNN, connecting to neurons in the upper layers to transform the data dimensions while preserving useful information. Their main function is to learn the output features of the preceding layers, perform global integration, and map them to the sample label space.

[0064] (1) A method for converting 16-dimensional features into a two-dimensional input matrix.

[0065] In this invention, the 16-dimensional feature vector needs to be reshaped into a two-dimensional matrix according to the principle of "process association grouping + time window expansion". The specific steps are as follows: 1. Process logic grouping: According to the reaction mechanism of the two-stage gasifier, the 16-dimensional feature is divided into 4 process association groups. Each group of features corresponds to the key link of the gasification process, avoiding meaningless spatial topology.

[0066] Group 1 (Coal Quality Basics): Elemental analysis of hydrogen mass fraction, elemental analysis of carbon mass fraction, moisture content in coal, and water-to-carbon ratio (4 characteristics that reflect the raw material basis for hydrogen generation).

[0067] Group 2 (Single-stage operation group): First-stage pulverized coal feed rate, second-stage pulverized coal feed rate, first-stage burner steam total volume, first-stage oxygen total flow rate (4 characteristics, controlling the reaction intensity of the first-stage gasification chamber).

[0068] Group 3 (Two-stage control group): Oxygen pressure, steam pressure, oxygen / steam pressure ratio, and carbon dioxide volume fraction (4 characteristics that affect the uniformity of the reaction in the two-stage gasification chamber).

[0069] Group 4 (Product Monitoring Group): Total syngas flow rate, methane volume fraction, hydrogen gas integral (actual value), carbon conversion rate (4 characteristics, feedback on the final state of the reaction).

[0070] 2. Time window expansion: Introduce a time window length T (in this embodiment, T=4, corresponding to the gasifier reaction delay period, determined by orthogonal experiment), take 4 sets of features from T consecutive sampling times, and stack them to form a two-dimensional matrix of "number of feature groups × time window length" (i.e., 4 groups × 4 times = 4 × 4 matrix).

[0071] 3. Input format adaptation: The final input tensor shape of the 2D CNN is [Batch_size, Group_num, Time_window, Channel], where Batch_size=64 (batch size), Group_num=4 (number of feature groups), Time_window=4 (time window), and Channel=1 (single channel, adapted for 2D convolution operation), i.e. [64,4,4,1].

[0072] (2) Convolutional neural network layer topology parameters.

[0073] This invention designs specific structural parameters for the CNN layer based on the characteristics of gasifier time-series data, configured as follows: 1. Convolutional layer: Number of convolutional kernels: 16; Convolutional kernel size: 3×3 (covering 3 time windows horizontally and 3 feature groups vertically to avoid extracting irrelevant features across process stages); Stride: Horizontal stride 1, vertical stride 1 (ensuring no loss of temporal continuity and process group association information); Padding method: Same Padding, ensuring that the output matrix and input matrix have the same size to maintain temporal integrity.

[0074] 2. Activation layer: The ReLU function is used to solve the gradient vanishing problem and adapt to the nonlinear characteristics of gasification data.

[0075] 3. Pooling layer: Pooling type: max pooling (extracts key temporal peak features, such as abrupt changes in operating parameters); pooling kernel size: 2×2 (compresses data dimensions while retaining core response information); stride: 2 (compresses the output matrix size to half of the input, i.e., 4×4→2×2).

[0076] 4. Dropout layer: Dropout rate 0.2 (to prevent overfitting, determined based on training set loss fluctuations).

[0077] 5. Flattening layer: The [64,2,2,16] tensor output after pooling is flattened into a one-dimensional vector (dimension = 2×2×16 = 64) and used as the input of the BiGRU layer.

[0078] Step S22: Construct a bidirectional gated recurrent unit (BiGRU) layer.

[0079] The gated recurrent unit (GRU) is a variant of the recurrent neural network (RNN), and its core mechanism revolves around two key gating mechanisms: the update gate and the reset gate.

[0080] like Figure 3 As shown, the update gate determines the proportion of hidden states from the past in the current state, affecting the continuity of information; the reset gate determines how the model integrates information based on the current input and the previous state, thereby adapting to the needs of dynamic changes in the sequence.

[0081] The GRU forward propagation process is shown in the following equation: ; ; ; ;in, To reset the door, For the current input, This is information implied in the previous moment. For the sigmoid function, [] indicates matrix stacking. To reset the weight of the door, To reset the door offset; To update the door, To update the weights of the gates, To update the door offset; Let be the candidate hidden state content, and tanh be the hyperbolic tangent function. The weights of the candidate hidden states, The bias for the candidate hidden state; It is in a hidden state.

[0082] As shown in Figure 4, the BiGRU of this invention integrates two GRU units: one processes information by forward propagation along the time series, and the other backtracks, thus enabling the model to simultaneously review the past and anticipate the future. The output of each BiGRU unit at time t is influenced by the bias at that time, the outputs of the forward GRU, and the backward GRU, as shown in the following formula: ; ; Where GRU(·) represents the calculation process of GRU. Let t be the output value of the GRU forward hidden layer. Let be the output value of the GRU backward hidden layer at time t, m be the output weight of the GRU forward hidden layer, and n be the output weight of the GRU backward hidden layer. for The corresponding hidden layer bias.

[0083] The BiGRU layer topology and input adaptation constructed in this invention are as follows: 1. Input adaptation logic: The 64-dimensional one-dimensional vector flattened by CNN is re-divided into time sequence according to the time window T=4, that is, each sample corresponds to 64-dimensional features of 4 time steps, forming the input sequence format of BiGRU [Batch_size, Time_window, Feature_dim] (i.e. [64,4,64]).

[0084] 2. BiGRU topology parameters: Number of hidden layer units: 64, with 32 each for forward GRU and backward GRU; Number of layers: 1, to avoid excessive model complexity leading to training instability; Activation function: tanh (adapts to the dynamic range of time-series data); Dropout: 0.2 (input layer dropout, consistent with CNN layers); Output method: Returns the hidden states at all time steps, dimension [64, 4, 64], for weighting by the Attention layer.

[0085] Step S23: Construct the attention mechanism layer.

[0086] Deep learning based on attention mechanisms aims to teach machines to distinguish between important and unimportant parts of data. Its structure is as follows: Figure 5 As shown, the calculation steps are as follows: 1) Encode the input sequence [y1, y2, ..., yn] to obtain a set of query vectors q; then, based on the similarity between scores [s1, s2, ..., sn], the scoring function is as follows: In the formula, , , All of these are matrices whose parameters are to be learned.

[0087] 2) Normalize using the Softmax function to obtain each input vector. The formula is as follows: .

[0088] 3) Calculate the weighted sum based on the weights and their corresponding value vectors to obtain the final hidden layer state value output 'a', as shown in the following formula: .

[0089] This invention describes the parameters of the Attention layer and the process correlation in the model constructed in this invention.

[0090] (1) Specific parameters of the Attention layer: Query vector (q): Dimension 64, consistent with the number of hidden units in the BiGRU hidden layer, obtained by mapping the hidden state of the last time step of BiGRU through a fully connected layer, as shown in the following formula: In the formula, It is a 64×64-dimensional learnable matrix. The bias is 64-dimensional, and the parameter dimensions match the feature complexity of the gasification data.

[0091] Key vector (k): 64 dimensions, obtained by mapping the hidden states of BiGRU across all time steps, as shown in the following formula: In the formula, It is a 64×64 dimensional matrix. 1, 2, 3, 4, corresponding to 4 time steps, are matched with the query vector dimensions to ensure the effectiveness of matrix operations.

[0092] (2) Process adaptation instructions for scoring and weighting.

[0093] The scoring function and weight normalization both employ the above general methods, the formulas of which have been described previously. The specific parameters in this invention's scenario are as follows: In the scoring function, the learnable vector... Set to 64 dimensions to adapt to the dimensions of query vectors and key vectors, avoiding gradient vanishing at high dimensions, and accurately capturing the nonlinear relationship of "abrupt changes in operating parameters → lagged changes in hydrogen content"; scoring results. Corresponding to the The contribution score of time step features to hydrogen prediction has a clear physical orientation.

[0094] Weight The distribution is strongly correlated with the gasifier's operating conditions: when the water-to-carbon ratio is in the optimal range (0.6~0.8) or the oxygen pressure is stable at a certain time step, Automatically increases to 0.6~0.7 to enhance the impact of key operating conditions; Softmax normalization ensures that the weights of the four time steps sum to 1, which conforms to the industrial logic of "cumulative contribution of multi-time operation to hydrogen generation" and avoids weight allocation from deviating from physical reality.

[0095] Step S24: Construct the output layer.

[0096] The context vector output by the attention mechanism layer is passed to a fully connected (Dense Layer), and finally outputs a scalar value through a linear activation function, which is the normalized predicted value of hydrogen content.

[0097] Step S25, embedding of physical constraints.

[0098] This invention incorporates the law of conservation of hydrogen mass in the gasification process as prior knowledge into the model. The hydrogen conservation equation is expressed as: Total hydrogen input to the gasifier = Total hydrogen output to the gasifier.

[0099] This is because in a two-stage dry pulverized coal gasifier, pulverized coal and gasifying agents (steam, oxygen) undergo complex physicochemical changes under high temperature and pressure, ultimately producing syngas with H2, CO, CO2, and CH4 as its main components. The migration and transformation of hydrogen follows the law of conservation of mass, and its core reaction mechanism is as follows:

[0100] (1) Pulverized coal pyrolysis and volatile matter reaction.

[0101] After the pulverized coal enters the first gasification chamber, it first undergoes a pyrolysis reaction, releasing volatiles (including H2, CH4, etc.). The reaction formula is as follows: .

[0102] (2) The reaction between coke and gasifying agent.

[0103] The coke produced by pyrolysis undergoes a gasification reaction with steam (H2O) and oxygen (O2) at high temperatures, generating H2, CO, and other gases. The core reactions include: ; ; ; .

[0104] (3) Methane formation reaction.

[0105] In the medium- and low-temperature range, some H2 will react with carbon or CO to produce CH4, consuming hydrogen and reducing the H2 content in the syngas. ; .

[0106] In summary, the migration of hydrogen within the gasifier follows the principle of "total input = total output," where the input hydrogen comes from the inherent hydrogen in the pulverized coal and the hydrogen in the steam; the output hydrogen exists in the syngas as H2 and CH4. All reactions do not change the total mass of hydrogen, only its form of existence.

[0107] Therefore, the specific "total input = total output" of hydrogen can be quantified as follows: ; 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 elemental analysis shows the hydrogen mass fraction (%). The total steam volume of a single burner (kg / h) Moisture content (%) in coal. This represents the total syngas flow rate (Nm³ / h). The predicted H2 volume fraction (%) is given by the model. The volume fraction of methane is (%).

[0108] Design the total loss function. From the data error term and mechanism constraint terms The composition is shown in the following formula: In the formula, These are the weighting coefficients for the physical constraint terms.

[0109] Specifically, the data error term The calculation formula is as follows: .

[0110] Physical constraints The calculation formula is as follows: ...

[0111] Combining the two, we get the following formula: In the formula, L represents the total loss. The H2 volume fraction predicted by the model for the i-th data point. The actual measured value of H2 volume fraction for the i-th data point. Let the total amount of hydrogen be the input for the i-th data point. The total output hydrogen amount for the i-th data point. is the weighting coefficient, and N is the number of data points.

[0112] Step S3: Model training and optimization.

[0113] The hybrid model constructed in S2 is trained using the training set prepared in S1. In this embodiment, the training parameters are set as follows: Optimizer: Adam; Initial learning rate: 0.005; Learning rate decay policy: decays every 150 epochs with a decay factor of 0.5; Batch size: 64; Number of epochs: 150; To prevent overfitting, a Dropout layer is added after the CNN and BiGRU layers, with a dropout rate set to 0.2.

[0114] This invention requires gradient descent optimization of the model and mechanistic parameters to update the parameters and minimize the total loss. The parameters are adjusted, and the above steps are repeated until the model converges to a suitable state.

[0115] The training process was completed on a computer equipped with an AMD Ryzen 5 3500U processor and 12GB of memory, using the MATLAB R2023b Deep Learning Toolbox. During training, the optimizer continuously updated the network weights and bias parameters by minimizing the total loss function L through the backpropagation algorithm.

[0116] Step S4: Model testing and performance evaluation.

[0117] After training, the model was evaluated using an independent test set (1728 samples). The test set data was input into the model to obtain predicted values, which were then inversely normalized to convert them back to practical engineering units. The following four commonly used metrics were used to evaluate the predictive performance: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²).

[0118] The specific formula is shown below: In the formula, , , represents the actual measured value, the predicted value, and the mean of the actual measured value, respectively, and n represents the total number of samples.

[0119] This invention inputs test set data into the optimized hybrid model, obtains prediction results, and uses an evaluation function to analyze and evaluate the prediction results, verifying the model's prediction performance under different operating conditions. To assess the usefulness of adding derived feature values ​​for predicting hydrogen at the gasifier outlet, the original feature values ​​and the combined feature values ​​after adding derived feature values ​​are input into the same model (CNN-BiGRU-Attention) for simulation experiments. The prediction results are shown below. Figure 6 and Figure 7 The evaluation index values ​​for each model are shown in Table 2.

[0120] Table 2: Comparison of Prediction Results After Inputting Different Feature Values .

[0121] To evaluate the superiority of the hybrid model in predicting hydrogen at the gasifier outlet, simulation experiments were conducted using the same dataset, employing both pure data-driven (CNN-BiGRU-Attention) and hybrid model simulations. The comparison results for each model are shown below. Figure 8 and Figure 9 The evaluation index values ​​for each model are shown in Table 3.

[0122] Table 3: Comparison of Prediction Results for Different Models .

[0123] from Figures 6-9It can be seen that in the hydrogen prediction model at the gasifier outlet, when the target value is relatively stable, the predictions of various models can closely follow the target value with relatively small fluctuations. However, the hybrid model performs better than the CNN-BiGRU-Attention model, and the addition of derived values ​​is better than the original feature values. When the hydrogen value changes significantly, the predictions of the hybrid 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, especially when the input is the original feature value.

[0124] Tables 2 and 3 also show that the hybrid model outperforms the CNN-BiGRU-Attention model, and the addition of derived features is even better than the original feature values. After adding derived feature values, the MAE is 0.030431, RMSE is 0.040331, and MAPE is 0.13403%, representing decreases of 24.46%, 18.21%, and 25.56% respectively compared to before adding the derived feature values. The R² after adding derived features is 96.762%, an improvement of 2.64% compared to before adding the derived feature values, demonstrating that the derived features effectively improve the model accuracy.

[0125] This demonstrates that the addition of mechanism-based derived feature values ​​in this invention improves the accuracy and stability of predictions. The hybrid model has a MAE of 0.021609, RMSE of 0.02604, and MAPE of 0.08689%, which are 24.35%, 26.21%, and 25.82% lower than the pure data-driven (CNN-BiGRU-Attention) model, respectively. The hybrid model has an R² of 99.456%, which is 2.32% higher than the pure data-driven (CNN-BiGRU-Attention) model. This indicates that physical constraints not only ensure the physical rationality of the results but also significantly improve prediction accuracy and stability. This is especially true during periods of drastic changes in operating conditions, such as... Figure 7 As shown, the prediction curve of the hybrid model is smoother, has a stronger ability to track dynamic changes, and has less lag and fluctuation.

[0126] Step S5, online application.

[0127] The trained optimal model is deployed to the online system. The real-time data acquisition module continuously obtains the raw variables listed in Table 1 from the DCS. After undergoing the same derived feature calculation, outlier handling, and normalization preprocessing as in the training phase, a 16-dimensional feature vector is generated. This vector is then input into the saved model to output the predicted value of hydrogen content at the gasifier outlet in real time, providing operators with key process indicators for optimizing gasifier operation.

[0128] The present invention also provides another embodiment of a two-stage gasifier outlet hydrogen dynamic prediction system, comprising: a data acquisition and preprocessing module for acquiring gasifier operating data in real time and completing derived feature construction, missing data processing, and data normalization; a feature extraction and temporal modeling module, comprising a convolutional neural network unit, a bidirectional gated recurrent unit, and an attention mechanism unit connected in sequence, for extracting features from the preprocessed data and modeling; a physical constraint calculation module for calculating the hydrogen mass conservation deviation based on the input data; and a model optimization and output module for optimizing model parameters based on a total loss function including data fitting loss and physical constraint loss, and outputting the final predicted hydrogen content value.

[0129] Preferably, the data acquisition and preprocessing module is specifically used to calculate at least one derived feature among the water-carbon ratio, oxygen-coal ratio, and oxygen / steam pressure ratio.

[0130] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and boxes Figure 1 The function specified in one or more boxes.

[0131] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A two-stage gasifier outlet hydrogen dynamic prediction method, characterized in that, The method comprises the following steps: Step S1, collecting historical operation data of the gasifier, constructing at least one derived feature data set based on the historical operation data, pre-processing the derived feature data set, and dividing the derived feature data set into a training set and a test set; the historical operation data comprises coal quality characteristic parameters and gasifier operation parameters; Step S2, constructing a deep learning model, the deep learning model comprising a convolutional neural network layer, a bidirectional gated recurrent unit layer and an attention mechanism layer in sequence; embedding a physical constraint term based on hydrogen element mass conservation in a loss function of the deep learning model to construct a total loss function; Step S3, training the deep learning model using the training set to obtain an optimized model; Step S4, inputting the test set into the optimized model to obtain a predicted verification value, and analyzing and evaluating the predicted verification value through an evaluation function constructed; Step S5, acquiring real-time operation data of the gasifier, inputting the real-time operation data into the optimized model, and outputting a real-time predicted value of hydrogen gas at the outlet of the gasifier; The derived features comprise at least one of a water-carbon ratio, an oxygen-coal ratio and an oxygen / steam pressure ratio; The water-carbon ratio is calculated from the total amount of steam in the first-stage burner, the first-stage coal powder feeding rate and the mass fraction of carbon in the coal powder; The oxygen-coal ratio is calculated from the total flow of oxygen in the first stage and the first-stage coal powder feeding rate; The oxygen / steam pressure ratio is calculated from the oxygen pressure and the steam pressure; The calculation formula of the water-carbon ratio is: ； wherein, represents the water carbon ratio, represents the steam flow rate of the first gasification chamber, represents the mass flow rate of the pulverized coal into the first gasification chamber, represents the mass fraction of carbon in the pulverized coal; The calculation formula of the oxygen-coal ratio is: ； wherein, represents the oxygen coal ratio, represents the oxygen volume flow of the first gasification chamber; The calculation formula of the oxygen / steam pressure ratio is: ； wherein, is the oxygen / steam pressure ratio, represents the pressure of oxygen into the burner, represents the pressure of steam into the burner.

2. The method of claim 1, wherein: In step S1, the pre-processing specifically comprises missing data processing, data dimension reduction and data normalization; wherein the missing data processing adopts 3σ-rule to detect abnormal data and replaces the abnormal data based on the change rate of adjacent time data.

3. The method of claim 1, wherein, In step S2: The physical constraint term based on hydrogen element mass conservation is calculated from the deviation of the total amount of hydrogen element input and the total amount of hydrogen element output; The total loss function is the weighted sum of a data fitting loss term and the physical constraint term.

4. The two-stage gasifier outlet hydrogen dynamic prediction method according to claim 3, characterized in that: The total amount of hydrogen element input is calculated based on the coal powder feeding rate, the hydrogen element content and the steam flow in the coal; The total amount of hydrogen element output is calculated based on the total flow of synthesis gas, the predicted hydrogen gas volume fraction and the methane volume fraction.

5. The two-stage gasifier outlet hydrogen dynamic prediction method according to claim 1, characterized in that: In step S1, the collected historical operation data comprises coal quality characteristic parameters, gasifier operation parameters and part of the detection parameters of the quenching system.

6. The two-stage gasifier outlet hydrogen dynamic prediction method according to claim 1, characterized in that: The convolutional neural network layer is used to extract local correlation features of the input parameters; The bidirectional gated recurrent unit layer is used to capture the dynamic time sequence law of the hydrogen content; The attention mechanism layer is used to strengthen the influence weight of key parameters on the prediction result.

7. A two-stage gasifier outlet hydrogen dynamic prediction system using the two-stage gasifier outlet hydrogen dynamic prediction method according to any one of claims 1 to 6, characterized by, comprise: a data acquisition and pre-processing module for acquiring real-time operation data of the gasifier, and completing derived feature construction, missing data processing and data normalization; The feature extraction and time sequence modeling module comprises a convolutional neural network unit, a bidirectional gated recurrent unit and an attention mechanism unit connected in sequence, and is used for extracting features and modeling from the preprocessed data; The physical constraint calculation module is used for calculating the hydrogen mass conservation deviation according to the input data; The model optimization and output module is used for optimizing the model parameters according to a total loss function comprising a data fitting loss and a physical constraint loss, and outputting a final hydrogen content prediction value.

8. The two-stage gasifier outlet hydrogen dynamic prediction system of claim 7, wherein: The data acquisition and preprocessing module is specifically used for calculating at least one derived feature in a water-carbon ratio, an oxygen-coal ratio and an oxygen / steam pressure ratio.

Images