Biomass gasification product distribution prediction method based on hard constraint physical information neural network

Abstract

The invention discloses a biomass gasification product distribution prediction method based on a hard constraint physical information neural network. The method comprises the following steps: collecting and preprocessing gasification experiment data; constructing a multi-layer artificial neural network model, and converting prior monotonicity knowledge into an inequality constraint combination only related to network parameters by adopting a hard constraint learning mode; model training: aiming at regression loss and regularization loss of experimental data, adopting a constrained particle swarm optimization algorithm to perform network parameter optimization on the model under an inequality constraint combination to obtain a hard constraint physical information neural network model; and predicting biomass gasification input data by adopting a hard constraint physical information neural network model to obtain a biomass gasification product distribution condition. According to the method, the monotonicity knowledge between the biomass gasification products and the key input parameters is embedded into the neural network model, the problem that biomass gasification experiment samples are insufficient can be effectively solved, and a more accurate and more interpretable biomass gasification product distribution result is obtained.

Description

Technical Field

[0001] The present invention relates to a gasification product distribution prediction technology, and in particular to a biomass gasification product distribution prediction method based on a hard-constrained physical information neural network. Background Art

[0002] With the steady growth of global energy demand and growing concerns about climate change, a diversified and sustainable energy production structure is of great significance. Biomass is one of the world's most abundant and economically viable renewable resources, offering a promising alternative to fossil fuels. Furthermore, biomass has the benefit of multiple utilization pathways and has the potential to support a continuous and reliable energy supply. Compared to direct combustion, gasification is a cost-effective and environmentally friendly biomass thermochemical conversion technology that can convert low-grade solid biomass into syngas, which is easy to store and transport.

[0003] However, improving gasification efficiency and syngas quality remains challenging. Syngas composition (e.g., hydrogen, carbon monoxide, etc.) is influenced by numerous factors, including feedstock characteristics (e.g., moisture content, particle size) and gasification conditions (e.g., temperature, equivalence ratio, catalyst usage). Traditional experimental research methods are costly and time-consuming, making it difficult to fully explore the underlying reaction mechanisms of the gasification process.

[0004] To address the shortcomings of experimental research, researchers have developed numerical models, which mainly fall into two categories: mechanical modeling and data-driven modeling. Although mechanical modeling methods (such as thermodynamic equilibrium models, kinetic models, and computational fluid dynamics models) can describe the physical and chemical phenomena of the gasification process, they suffer from problems such as difficulty in obtaining parameters, high computational costs, and insufficient prediction accuracy. Data-driven modeling methods, namely machine learning technology, predict gasification performance by learning from experimental data. It has the advantages of low modeling cost and high accuracy, but it also suffers from strong data dependence, poor physical consistency, and insufficient interpretability.

[0005] To address these issues, researchers have proposed a physical-informed machine learning approach, incorporating prior physical knowledge into machine learning models to improve their interpretability and generalization capabilities. However, current machine learning-based biomass gasification models generally face a core bottleneck: insufficient generalization across feedstocks and operating conditions. Traditional methods are prone to deviations from mechanistic principles and a sharp drop in accuracy in prediction scenarios outside the distribution range of the training set. Summary of the Invention

[0006] Purpose of the invention: The purpose of the present invention is to provide a method for predicting the distribution of biomass gasification products based on a hard-constrained physical information neural network. Under extreme working conditions where experimental data are sparse or not covered, the method can still strictly follow the gasification reaction laws and provide reliable predictions, thereby significantly improving the generalization ability across raw materials and working conditions.

[0007] Technical solution: The present invention provides a method for predicting the distribution of biomass gasification products based on a hard-constrained physical information neural network, comprising:

[0008] Collect gasification experimental data of different biomass raw materials under different operating conditions and preprocess the gasification experimental data;

[0009] Construct a multi-layer artificial neural network model and use hard constraint learning to transform the prior monotonicity knowledge into a combination of inequality constraints involving only the parameters of the multi-layer artificial neural network model.

[0010] A basic hard-constrained physical information neural network model was constructed. The preprocessed gasification experimental data was input into a multi-layer artificial neural network model and the model was trained. During the training process, the constrained particle swarm optimization algorithm was used to optimize the network parameters of the multi-layer artificial neural network model under the combination of inequality constraints, targeting the regression loss and regularization loss of the experimental data, and the hard-constrained physical information neural network model was obtained.

[0011] The hard-constrained physical information neural network model is used to predict the biomass gasification input data and obtain the distribution of biomass gasification products.

[0012] Furthermore, the gasification experimental data includes input variables and output variables, wherein the input variables include temperature, humidity, and equivalence ratio, and the output variables include hydrogen, carbon monoxide, carbon dioxide, methane, and nitrogen concentrations.

[0013] Furthermore, the gasification experimental data were preprocessed in the following way: when the categorical variable had only two values, binary label encoding was used; when the categorical variable contained more than two categories, one-hot encoding was used to convert the categorical variable into a multidimensional binary vector.

[0014] Furthermore, a multi-layer artificial neural network model is constructed, and a hard constraint learning method is used to transform the prior monotonicity knowledge into a combination of inequality constraints involving only the parameters of the multi-layer artificial neural network model, including:

[0015] Extract the physical monotonic relationship between syngas composition and key operating parameters in biomass gasification process;

[0016] Construct a multi-layer artificial neural network model;

[0017] The Sigmoid activation function is used to transform the physical monotonicity relationship into a hard constraint condition that only contains the parameters of the multi-layer artificial neural network model.

[0018] Furthermore, the multi-layer artificial neural network model includes an input layer, N hidden layers and an output layer;

[0019] Among them, the input layer uses a multi-dimensional vector as the input of the multi-layer artificial neural network model. The input vector x is expressed as follows:

[0020]

[0021] Where n is the number of input parameters;

[0022] The hidden layer contains neural nodes, and the number of neural nodes in the i-th hidden layer is represented by h (i) ,1≤i≤N;

[0023] The output layer uses a multi-dimensional vector as the output of the multi-layer artificial neural network model. The output vector y is expressed as follows:

[0024]

[0025] Where c is the number of output parameters.

[0026] Furthermore, the calculation formula of the model output of the multi-layer artificial neural network model is as follows:

[0027]

[0028] in, Represents the predicted value of the multi-layer artificial neural network model, β represents the weight matrix connecting the Nth hidden layer and the output layer, Sigmoid function is the activation function; w (N) and b (N) Represent the weight matrix and bias vector of the Nth hidden layer respectively, x (N) represents the input vector of the Nth hidden layer,

[0029] x (N) It can be given by the following formula:

[0030] x (N) =g(w (N-1) x (N-1) +b (N-1) )

[0031] Among them, w (N-1) 、b (N-1) , and x (N-1) , respectively represent the weight matrix, bias vector and input vector of the N-1th hidden layer, and The input vector of the first hidden layer is the input vector x of the multi-layer artificial neural network model.

[0032] Furthermore, the objective function of the multi-layer artificial neural network model is:

[0033]

[0034] in, represents the regression loss; y (j) represents the jth sample; represents the predicted value of the jth sample; m represents the number of training samples; ‖·‖2 represents the two-norm; β represents the weight matrix connecting the Nth hidden layer and the output layer; w (i) and b (i) denote the weight matrix and bias vector of the i-th hidden layer respectively.

[0035] Furthermore, the Sigmoid activation function is used to transform the physical monotonicity relationship into hard constraints that only contain the parameters of the multi-layer artificial neural network model, including:

[0036] The physical monotonicity relationship can be expressed by the objective function of the multi-layer artificial neural network model as follows:

[0037]

[0038] Among them, x k Represents the kth input variable, the kth input variable and the lth prediction target y l There is a monotonic relationship; represents a positive monotonic relationship, Indicates a negative monotonic relationship;

[0039] According to the Sigmoid activation function, there is 0 <g (i)′ <1, 1≤i≤N, where g (i)′ Represents the derivative of the activation function of the i-th hidden layer. The multi-layer artificial neural network model includes N hidden layers, and the model output is Regarding the kth input variable x k The partial derivatives of are expressed as follows:

[0040]

[0041] Therefore, the positive monotonicity relationship Can be replaced by no input variable constraints

[0042]

[0043] in, Represents model output Regarding the kth input variable x k The dth constraint under the general monotonicity constraint of ;

[0044] Negative monotonicity relationship Can be replaced by

[0045]

[0046] Furthermore, based on the regression loss and regularization loss of the experimental data, the constrained particle swarm optimization algorithm is used to optimize the network parameters of the multi-layer artificial neural network model under the combination of inequality constraints, and a hard-constrained physical information neural network model is obtained, including:

[0047] The particle swarm optimization algorithm uses the velocity vector to update the current position of each particle in the swarm until the global optimal particle in the feasible region is identified, at which point the search terminates; the position of the oth particle at iteration z+1 is as follows:

[0048]

[0049] Where, represents the corresponding update velocity vector; Δt represents the unit time step; u represents the inertia weight; c1 and c2 represent cognitive parameters and social parameters respectively; r1 represents the random weight that controls the movement of particles to their own historical optimal positions, which is a random number between 0 and 1; r2 represents the random intensity that regulates the particles to approach the global optimal position of the group, which is a random number between 0 and 1; and They represent the optimal position of the oth particle and the global optimal position before iteration z times respectively;

[0050] The constraints are adapted using a parameter-free adaptive penalty scheme as follows:

[0051]

[0052] Where Γ(p) and Represent the objective function and its average value in the current group; b d represents the penalty parameter; represents the average violation of the d-th constraint in the current population.

[0053] At the beginning of the training process, the input and output data are normalized based on Min-Max to eliminate the effect of variable scale;

[0054] Introducing regularization loss terms Prevent overfitting by penalizing network parameters; loss function of hard-constrained physical information neural network It can be expressed as follows:

[0055]

[0056] Where λ reg represents the weight controlling the regularization.

[0057] Furthermore, the calculation formula used in the Min-Max based normalization process is as follows:

[0058]

[0059] Where, X norm represents normalized data; X max and X min The vectors representing the maximum and minimum values of each variable in the training sample set.

[0060] Beneficial effects: Compared with the prior art, the present invention has the following significant technical effects:

[0061] By converting the physically monotonic relationships between gasification temperature, equivalence ratio, and feedstock moisture content and syngas composition into hard constraints on neural network parameters, this approach directly embeds thermodynamic mechanism knowledge to achieve global physical consistency. This approach simultaneously optimizes model accuracy and mechanistic conformance through a constrained particle swarm optimization algorithm. Even under extreme conditions where experimental data is sparse or uncovered, it can strictly adhere to gasification reaction laws and provide reliable predictions, significantly improving generalization across feedstocks and operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a flow chart of the present invention;

[0063] Figure 2 This is a training framework diagram for the basic hard-constrained physical information neural network model in the present invention. DETAILED DESCRIPTION

[0064] The technical solution of the present invention is described in detail below in conjunction with specific implementation methods and the accompanying drawings.

[0065] like Figure 1 As shown, a method for predicting the distribution of biomass gasification products based on a hard-constrained physical information neural network of the present invention comprises the following steps:

[0066] S1. Collect gasification experimental data of different biomass raw materials under different operating conditions and preprocess the gasification experimental data.

[0067] The gasification experimental data include input variables (such as temperature, humidity, and equivalence ratio) and output variables (such as hydrogen, carbon monoxide, carbon dioxide, methane, and nitrogen concentrations).

[0068] In this embodiment, the gasification experimental data are preprocessed in the following manner: when the categorical variable has only two values, binary label encoding is used; when the categorical variable contains more than two categories, one-hot encoding is used to convert it into a multidimensional binary vector.

[0069] S2. Use hard constraint learning to transform the prior monotonicity knowledge into a combination of inequality constraints involving only network parameters. Step S2 specifically includes the following:

[0070] S2.1. The physical monotonic relationship between the syngas composition and key operating parameters during the extraction of biomass gasification is shown in the following table.

[0071]

[0072] S2.2. Construct a multi-layer artificial neural network model.

[0073] like Figure 2 As shown in Figure 1, the multi-layer artificial neural network model includes an input layer, N hidden layers, and an output layer. The input layer uses a multi-dimensional vector as the input of the multi-layer artificial neural network model. The input vector x is expressed as follows:

[0074]

[0075] Where n is the number of input parameters.

[0076] The hidden layer contains neural nodes, and the number of neural nodes in the i-th hidden layer is represented by h (i) ,1≤i≤N.

[0077] The output layer uses a multi-dimensional vector as the output of the multi-layer artificial neural network model. The output vector y is expressed as follows:

[0078]

[0079] Where c is the number of output parameters.

[0080] The calculation formula of the model output of the multi-layer artificial neural network model is as follows:

[0081]

[0082] in, Represents the predicted value of the multi-layer artificial neural network model, β represents the weight matrix connecting the Nth hidden layer and the output layer, Sigmoid function is the activation function; w (N) and b (N) Represent the weight matrix and bias vector of the Nth hidden layer respectively, x (N) represents the input vector of the Nth hidden layer,

[0083] x (N) It can be given by the following formula:

[0084] x (N) =g(w (N-1) x (N-1) +b (N-1) ) (4)

[0085] Among them, w (N-1) 、b (N-1) , and x (N-1) , respectively represent the weight matrix, bias vector and input vector of the N-1th hidden layer, and The input vector of the first hidden layer is the input vector x of the multi-layer artificial neural network model.

[0086] The objective function of the multi-layer artificial neural network model is:

[0087]

[0088] in, represents the regression loss; y (j) represents the jth sample; represents the predicted value of the jth sample; m represents the number of training samples; ‖·‖2 represents the two-norm; β represents the weight matrix connecting the Nth hidden layer and the output layer; w (i) and b (i) denote the weight matrix and bias vector of the i-th hidden layer respectively.

[0089] S2.3. Use the Sigmoid activation function to transform the physical monotonicity relationship into a hard constraint condition that only contains the parameters of the multi-layer artificial neural network model. The details are as follows:

[0090] The physical monotonicity relationship can be expressed by the objective function of the multi-layer artificial neural network model as follows:

[0091]

[0092] Among them, x k Represents the kth input variable, the kth input variable and the lth prediction target y l There is a monotonic relationship; represents a positive monotonic relationship, Indicates a negative monotonic relationship. Model output Regarding the kth input variable x kThe partial derivative of can be expanded as follows based on the chain rule and the characteristics of the Sigmoid activation function in neural networks:

[0093]

[0094] Among them, β l represents the lth row of matrix β, and β l,iN is the lth row and i-th row of the matrix β N Elements of the column; is the vector b (N) The i N Quantity; Represents the matrix w (N) The i N OK; and, is the matrix w (N) The i N row and i N-1 Elements of the column; and They are vector x (N) and b (N-1) The i N-1 Quantity; Represents the matrix w (N-1) The i N-1 The derivative of the activation function of the i-th hidden layer is expressed as g (i)′ Therefore, formula (7) can be simplified to:

[0095]

[0096] Sigmoid activation function has 0 <g (i)′ <1,1≤i≤N. Therefore, equation (8) can be simplified to its sufficient and unnecessary conditions:

[0097]

[0098] Therefore, the positive monotonicity relationship Can be replaced by no input variable constraints

[0099]

[0100] in, Represents model output Regarding the kth input variable x k The dth constraint under the general monotonicity constraint of .

[0101] Negative monotonicity relationship Can be replaced by

[0102]

[0103] Equations (11) and (12) are combinations of inequality constraints involving only network parameters.

[0104] S3. Input the preprocessed gasification experimental data into the multi-layer artificial neural network model to train the model; during the training process, based on the regression loss and regularization loss of the experimental data, the constrained particle swarm optimization algorithm is used to optimize the network parameters of the multi-layer artificial neural network model under the combination of inequality constraints to obtain a hard-constrained physical information neural network model.

[0105] During the training process, the constrained particle swarm optimization algorithm is used to optimize the network parameters of the multi-layer artificial neural network model under the combination of inequality constraints, aiming at the regression loss and regularization loss of the experimental data, and a hard-constrained physical information neural network model is obtained. Figure 2 As shown, the specific process is as follows:

[0106] S3.1, the particle swarm optimization algorithm uses the velocity vector to update the current position of each particle in the swarm until the global optimal particle in the feasible region is identified, and the search is terminated at the feasible region. The position of the oth particle at iteration z+1 as follows:

[0107]

[0108] Where, represents the corresponding update velocity vector; Δt represents the unit time step; u represents the inertia weight; c1 and c2 represent cognitive parameters and social parameters respectively; r1 represents the random weight that controls the movement of particles to their own historical optimal positions, which is a random number between 0 and 1; r2 represents the random intensity that adjusts the particles to approach the global optimal position of the group, which is a random number between 0 and 1. and They represent the optimal position of the oth particle before iteration z and the global optimal position respectively.

[0109] Furthermore, a parameter-free adaptive penalty scheme is used to adapt the constraints as follows:

[0110]

[0111] Where Γ(p) and Represent the objective function and its average value in the current group; b d represents the penalty parameter; represents the average violation of the d-th constraint in the current population.

[0112] S3.2. At the beginning of the training process, the input and output data are normalized based on Min-Max to eliminate the effect of variable scale. The calculation formula used for Min-Max normalization is as follows:

[0113]

[0114] Where, X norm represents normalized data; X max and X min The vectors representing the maximum and minimum values of each variable in the training sample set.

[0115] S3.3. Introducing regularization loss terms Prevent overfitting by penalizing network parameters. Loss function of hard-constrained physical information neural network It can be expressed as follows:

[0116]

[0117] Where λ reg represents the weight controlling the regularization.

[0118] S4. Use the hard-constrained physical information neural network model to predict the biomass gasification input data and obtain the distribution of biomass gasification products.

[0119] The present invention embeds the monotonicity knowledge between biomass gasification products and key input parameters into the neural network model, which can effectively solve the problem of insufficient biomass gasification experimental samples, obtain more accurate and interpretable biomass gasification product distribution results, and provide operators and designers with suggestions for operation optimization and process design.

Claims

1. A method for predicting biomass gasification product distribution based on a hard-constrained physical information neural network, characterized in that: include: Collect gasification experimental data of different biomass raw materials under different operating conditions and preprocess the gasification experimental data; A multi-layer artificial neural network model is constructed, and hard constraint learning is used to transform the prior monotonicity knowledge into a combination of inequality constraints involving only the network parameters of the multi-layer artificial neural network model. The preprocessed gasification experimental data was input into a multi-layer artificial neural network model for model training. During the training process, the constrained particle swarm optimization algorithm was used to optimize the network parameters of the multi-layer artificial neural network model under the combination of inequality constraints, targeting the regression loss and regularization loss of the experimental data. This resulted in a hard-constrained physical information neural network model. The hard-constrained physical information neural network model is used to predict the biomass gasification input data and obtain the distribution of biomass gasification products.

2. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 1 is characterized in that: The gasification experimental data includes input variables and output variables, wherein the input variables include temperature, humidity, and equivalence ratio, and the output variables include the concentrations of hydrogen, carbon monoxide, carbon dioxide, methane, and nitrogen.

3. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 1 is characterized in that: The gasification experimental data were preprocessed in the following way: when the categorical variable had only two values, binary label encoding was used; when the categorical variable contained more than two categories, one-hot encoding was used to convert the categorical variable into a multidimensional binary vector.

4. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 1 is characterized in that: A multi-layer artificial neural network model is constructed, and hard constraint learning is used to transform the prior monotonicity knowledge into a combination of inequality constraints involving only the network parameters of the multi-layer artificial neural network model, including: Extract the physical monotonic relationship between syngas composition and key operating parameters in biomass gasification process; Construct a multi-layer artificial neural network model; The Sigmoid activation function is used to transform the physical monotonicity relationship into a hard constraint condition that only contains the parameters of the multi-layer artificial neural network model.

5. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 4 is characterized in that: The multi-layer artificial neural network model includes an input layer, N hidden layers and an output layer; Among them, the input layer uses a multi-dimensional vector as the input of the multi-layer artificial neural network model. The input vector x is expressed as follows: Where n is the number of input parameters; The hidden layer contains neural nodes, and the number of neural nodes in the i-th hidden layer is represented by h (i) ,1≤i≤N; The output layer uses a multi-dimensional vector as the output of the multi-layer artificial neural network model. The output vector y is expressed as follows: Where c is the number of output parameters.

6. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 5, characterized in that: The calculation formula of the model output of the multi-layer artificial neural network model is as follows: in, Represents the predicted value of the multi-layer artificial neural network model, β represents the weight matrix connecting the Nth hidden layer and the output layer, Sigmoid function is the activation function; w (N) and b (N) Represent the weight matrix and bias vector of the Nth hidden layer respectively, x (N) represents the input vector of the Nth hidden layer, x (N) It can be given by the following formula: x (N) =g(w (N-1) x (N-1) +b (N-1) ) Among them, w (N-1) 、b (N-1) , and x (N-1) , respectively represent the weight matrix, bias vector and input vector of the N-1th hidden layer, and The input vector of the first hidden layer is the input vector x of the multi-layer artificial neural network model.

7. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 4, characterized in that: The objective function of the multi-layer artificial neural network model is: in, represents the regression loss; y (j) represents the jth sample; represents the predicted value of the jth sample; m represents the number of training samples; ‖·‖2 represents the two-norm; β represents the weight matrix connecting the Nth hidden layer and the output layer; w (i) and b (i) denote the weight matrix and bias vector of the i-th hidden layer respectively.

8. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 4, characterized in that: The Sigmoid activation function is used to transform the physical monotonicity relationship into hard constraints that only contain the parameters of the multi-layer artificial neural network model, including: The physical monotonicity relationship can be expressed by the objective function of the multi-layer artificial neural network model as follows: Among them, x k Represents the kth input variable, the kth input variable and the lth prediction target y l There is a monotonic relationship; represents a positive monotonic relationship, Indicates a negative monotonic relationship; According to the Sigmoid activation function, there is 0 <g (i)′ <1, 1≤i≤N, where g (i)′ Represents the derivative of the activation function of the i-th hidden layer. The multi-layer artificial neural network model includes N hidden layers, and the model output is Regarding the kth input variable x k The partial derivatives of are expressed as follows: Therefore, the positive monotonicity relationship Can be replaced by no input variable constraints in, Represents model output Regarding the kth input variable x k The dth constraint under the general monotonicity constraint of ; Negative monotonicity relationship Can be replaced by 9. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 1, characterized in that: During the training process, the constrained particle swarm optimization algorithm is used to optimize the network parameters of the multi-layer artificial neural network model under the combination of inequality constraints, targeting the regression loss and regularization loss of the experimental data. The hard-constrained physical information neural network model is obtained, including: The particle swarm optimization algorithm uses the velocity vector to update the current position of each particle in the swarm until the global optimal particle in the feasible region is identified, at which point the search terminates; the position of the oth particle at iteration z+1 is as follows: Where, represents the corresponding update velocity vector; Δt represents the unit time step; u represents the inertia weight; c1 and c2 represent cognitive parameters and social parameters respectively; r1 represents the random weight that controls the movement of particles to their own historical optimal positions, which is a random number between 0 and 1; r2 represents the random intensity that regulates the particles to approach the global optimal position of the group, which is a random number between 0 and 1; and They represent the optimal position of the oth particle and the global optimal position before iteration z times respectively; The constraints are adapted using a parameter-free adaptive penalty scheme as follows: Where Γ(p) and Represent the objective function and its average value in the current group; b d represents the penalty parameter; represents the average violation of the d-th constraint in the current population. At the beginning of the training process, the input and output data are normalized based on Min-Max to eliminate the effect of variable scale; Introducing regularization loss terms Prevent overfitting by penalizing network parameters; loss function of hard-constrained physical information neural network It can be expressed as follows: Where λ reg represents the weight controlling the regularization.

10. The method for predicting biomass gasification product distribution based on hard-constrained physical information neural network according to claim 9, characterized in that: The calculation formula used in the Min-Max-based normalization process is as follows: Where, X norm represents normalized data; X max and X min The vectors representing the maximum and minimum values of each variable in the training sample set.

Images