Biomass gasification reaction modeling method and system based on physical information neural network

Abstract

The present application relates to the technical field of biomass energy process modeling, and discloses a biomass gasification reaction modeling method and system based on a physical information neural network.The method comprises the following steps: obtaining original gasification data and performing data tensorization to obtain a characteristic tensor; training an adversarial network model to generate an enhanced gasification characteristic data set covering unseen working conditions; constructing a residual double-flow network architecture including a thermodynamic equilibrium subnetwork for calculating a theoretical benchmark and a non-equilibrium dynamics subnetwork for learning deviation correction; constructing a physical constraint loss function to fine-tune the residual double-flow network architecture, thereby obtaining a physical constraint optimization model; and performing reverse optimization on the physical constraint optimization model to output optimized operation parameters.The present application solves the problems of prediction deviation and violation of physical laws of traditional models, realizes a functional closed loop from high-precision prediction to active process optimization, and significantly improves the prediction accuracy and generalization ability of the model under raw material fluctuations.

Description

Technical Field

[0001] This invention relates to the field of biomass energy process modeling technology, and more specifically, to a biomass gasification reaction modeling method and system based on physical information neural networks. Background Technology

[0002] As biomass gasification technology plays an increasingly important role in energy transition, the predictive accuracy bottleneck of its process control models has become a key constraint on industrial development. Traditional models exhibit significant deviations between predicted and actual values ​​when dealing with fluctuations in feedstock or changes in operating conditions, severely impacting the stability of downstream processes. Existing technologies suffer from two major limitations: thermodynamic equilibrium-based mechanistic models neglect significant non-equilibrium kinetic effects during gasification, such as tar formation and incomplete carbon conversion, leading to severe prediction distortion; while purely data-driven machine learning models, when faced with sparse, high-dimensional data from industrial settings, are prone to overfitting and often output absurd results that violate fundamental physical laws such as the conservation of mass and energy. Traditional hybrid modeling methods often remain at the level of simple linear correction, failing to achieve a deep integration of mechanisms and data. Therefore, how to combine explicit physical laws as hard constraints with data-driven nonlinear fitting capabilities to construct a modeling architecture that both adheres to physical laws and adaptively corrects non-equilibrium deviations, enabling high-precision and interpretable predictions under high-dimensional, sparse data conditions, is a pressing technical challenge in this field.

[0003] In the prior art, Chinese patent CN117008479B discloses a method and system for optimizing and controlling negative carbon emissions based on a biomass gasifier. This system includes a data preprocessing unit, a parameter feature classification module, a distribution feature extraction unit, a strategy network calculation module, and a gasifier control algorithm module, enabling functions such as gasification data classification, distribution feature extraction, dynamic model construction, and operational parameter optimization. It constructs a dynamic analysis model of the gasification reaction using an LSTM network and a fully connected network, improving the accuracy of biomass gasification parameter prediction and providing data support for adjusting gasifier operating parameters. Chinese patent application CN120409193A discloses a twin optimization system and method for the biomass gasification process. This system adopts a multi-module collaborative architecture, including modules for multi-source data intelligent sensing, hybrid evolutionary algorithm self-optimization, knowledge graph adaptive reconstruction, multi-timescale collaborative calibration and prediction, and closed-loop feedback iterative symbiosis, respectively realizing model bias identification, parameter optimization, model reconstruction, accurate prediction, and iterative optimization functions. By combining digital twin technology with evolutionary algorithms, it enables the gasification process to shift from passive early warning to active optimization, thereby improving the long-term consistency and adaptability of the model.

[0004] However, while the two existing technologies mentioned above have some value in gasification data processing, model optimization, and process control, they fail to address the core pain point of synergistic coordination between physical law constraints and adaptive correction of non-equilibrium deviations in current biomass gasification modeling. Specifically, the patent with authorization announcement number CN117008479B focuses on data-driven dynamic prediction, failing to integrate physical laws such as mass and energy conservation as hard constraints into the model architecture. This easily leads to results that violate basic physical laws and lacks anti-overfitting design when dealing with sparse, high-dimensional data. The patent with publication number CN120409193A focuses on the combination of digital twins and algorithm optimization, failing to achieve a deep integration of mechanistic knowledge and data-driven capabilities. Its correction effect on non-equilibrium kinetic effects such as tar formation and incomplete carbon conversion during gasification is limited. Neither technology constructs a modeling architecture that coordinates physical constraints and nonlinear fitting. This makes it impossible to maintain prediction accuracy under fluctuating feedstocks or changing operating conditions, and it is also difficult to balance model interpretability and adaptability, failing to meet the biomass gasification industry's demand for high-precision, high-reliability process control models. Summary of the Invention

[0005] This invention is applicable to various biomass gasification processes, such as fluidized bed or fixed bed gasifiers, and can meet the dynamic modeling needs of biomass feedstocks from different sources. By constructing raw gasification data including experimental and simulation log files and performing data tensor operations, a standardized data foundation is provided. An adversarial network model is used to enhance the feature tensors across domains, generating an enhanced gasification feature dataset covering a wider range of operating conditions, solving the problem of sparsity in real experimental data and systematic bias in simulation data. A residual dual-flow network architecture is constructed, using a thermodynamic equilibrium flow subnet to provide a theoretical benchmark and a non-equilibrium dynamic flow subnet to learn bias corrections, decomposing complex nonlinear predictions into a "benchmark + correction" problem, significantly improving prediction accuracy. By constructing a physical constraint loss function containing conservation laws and dynamic monotonicity constraints, the residual dual-flow network architecture is fine-tuned and trained, and the resulting physical constraint optimization model ensures the physical consistency of the prediction results. Finally, reverse optimization is performed, and the output optimized operating parameters realize a functional closed loop from passive prediction to active process optimization, providing core technical means for achieving intelligent and optimized operation of the biomass gasification process.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A biomass gasification reaction modeling method based on physical information neural networks includes:

[0008] The raw gasification data characterizing the biomass gasification reaction process is obtained. Data tensor operation is performed on the raw gasification data to output the feature tensor for data augmentation. Adversarial game training is performed on the feature tensor to obtain the adversarial network model. The working condition area not covered by the experiment is sampled and the adversarial network model is called to obtain the enhanced gasification feature dataset.

[0009] Thermodynamic equilibrium flow subnet calculations are performed based on the enhanced gasification feature dataset to obtain a theoretical equilibrium prediction result set. A non-equilibrium dynamic flow subnet is constructed and trained using the theoretical equilibrium prediction result set. The non-equilibrium dynamic flow subnet is called for calculation to obtain the predicted dynamic deviation vector. The predicted dynamic deviation vector is fused to obtain the final syngas component prediction result set, thus completing the pre-construction of the residual two-flow network architecture.

[0010] A physical constraint loss function is constructed to fine-tune the residual two-stream network architecture, resulting in a physical constraint optimization model. Inverse optimization is then performed on the physical constraint optimization model to output optimized operating parameters for guiding the gasification process.

[0011] Furthermore, the method for obtaining the adversarial network model includes:

[0012] The feature tensor includes experimental source feature tensors derived from experimental log files and simulation source feature tensors derived from simulation log files. The feature tensor is obtained by concatenating the raw material feature vector and the operating condition feature vector, which are obtained by performing normalization processing on the original gasification data.

[0013] Construct an adversarial network model that includes a condition generator and a convolutional discriminator, wherein the network structure of the condition generator is a deep residual network and the network structure of the convolutional discriminator is a multi-layer convolutional neural network.

[0014] After constructing the condition generator and the convolutional discriminator, adversarial game training is performed. This training alternately optimizes the convolutional discriminator and the condition generator until the total loss function converges to below the preset adversarial threshold, thereby outputting the adversarial network model.

[0015] Furthermore, the optimization of the convolutional discriminator includes:

[0016] Treating the network weight parameters inside the condition generator as constants, a convolutional discriminator loss function is calculated. The convolutional discriminator loss function is obtained by performing algebraic operations on the scalar values ​​corresponding to the virtual high-fidelity feature tensor generated by the condition generator and the scalar values ​​corresponding to the experimental source feature tensor.

[0017] The partial derivatives of the convolutional discriminator loss function with respect to the internal network weight parameters of the convolutional discriminator are calculated using the backpropagation algorithm, and the weights are updated using an optimizer to minimize the convolutional discriminator loss function.

[0018] Furthermore, the method for obtaining the theoretical equilibrium prediction result set includes:

[0019] The enhanced gasification feature dataset includes simulated source feature tensors, experimental source feature tensors, and enhanced feature tensors for regions not covered by experiments;

[0020] For each feature tensor in the enhanced gasification feature dataset, atomic composition information, gasification temperature, and gasification pressure are extracted from the working condition feature vector and raw material feature vector contained therein as input parameters.

[0021] The input parameters are fed into a thermodynamic equilibrium subnet for calculation, resulting in a set of solutions describing the molar number of each gas phase product under equilibrium conditions.

[0022] The solutions are converted into their respective volume fractions in the total gas to obtain the theoretical equilibrium prediction vector. The total gas refers to the mixture of all products in the reaction system that exist in gaseous form when thermodynamic equilibrium is reached. The theoretical equilibrium prediction vectors corresponding to each feature tensor in the enhanced gasification feature dataset are combined to obtain the theoretical equilibrium prediction result set.

[0023] Furthermore, the non-equilibrium dynamic subnet includes:

[0024] For each feature tensor in the enhanced gasification feature dataset, extract the corresponding value of the feature tensor from the corresponding original gasification data record to form a target state vector. Subtract the target state vector from the theoretical equilibrium prediction vector corresponding to the feature tensor element by element to obtain a target deviation vector. Combine the target deviation vectors corresponding to each feature tensor to obtain the target deviation dataset.

[0025] The enhanced gasification feature dataset and the corresponding target deviation dataset are divided into a training set and a validation set. In each training iteration cycle, a training batch is extracted from the training set, and the joint input vector in the training batch is input into the non-equilibrium dynamic flow subnet to obtain the prediction deviation batch. The mean square error between the prediction deviation batch and the target deviation vector is calculated as the loss function, and the network weights are updated through backpropagation until the loss function converges on the validation set to obtain the non-equilibrium dynamic flow subnet.

[0026] Furthermore, the method for obtaining the final syngas component prediction result set includes:

[0027] For each feature tensor in the enhanced gasification feature dataset, the non-equilibrium dynamics subnet is invoked to calculate the prediction dynamics bias vector corresponding to the feature tensor;

[0028] The predicted dynamics deviation vector and the theoretical equilibrium state prediction vector corresponding to the feature tensor are added element by element to obtain a syngas component prediction vector. Each element value in the syngas component prediction vector is the result of adding the corresponding element value in the theoretical equilibrium state prediction vector and the corresponding element value in the predicted dynamics deviation vector.

[0029] The syngas component prediction vectors calculated from all feature tensors are combined to obtain the final syngas component prediction result set.

[0030] Furthermore, the physical constraint loss function includes:

[0031] The physical constraint loss function consists of two parts. The first part is the conservation law constraint term, which includes the element mass conservation residual and the system enthalpy change residual. The element mass conservation residual is obtained by calculating the absolute value of the difference between the initial total element mass of the reaction system and the predicted total element mass of the product. The initial total element mass of the reaction system is calculated based on the raw material feature vector and the operating condition feature vector in the feature tensor. The predicted total element mass of the product is calculated based on the residual two-stream network architecture. The system enthalpy change residual is obtained by calling a thermodynamic database.

[0032] The second part is the kinetic monotonicity constraint term, which is obtained by calculating the partial derivative of the syngas component prediction vector with respect to the input gasification temperature, and generating a positive penalty value when the partial derivative is negative.

[0033] The physical constraint loss function is obtained by weighted summation of the conservation law constraint term and the dynamic monotonicity constraint term.

[0034] Furthermore, the method for obtaining the physical constraint optimization model includes:

[0035] A composite loss function is constructed by weighted summation of two components. The first part is the data fitting error term obtained by calculating the mean square error between the syngas component prediction vector output by the residual two-stream network architecture and the actual output result corresponding to the feature tensor. The second part is the physical constraint loss function.

[0036] Using the composite loss function, the partial derivatives of the composite loss function with respect to the network weight parameters inside the non-equilibrium dynamic flow subnet are calculated through the backpropagation algorithm. The weights are then updated using an optimizer until the value of the composite loss function converges to below a preset fine-tuning convergence threshold, thus obtaining the physical constraint optimization model.

[0037] Furthermore, the method for obtaining the optimized operating parameters includes:

[0038] Define a process optimization objective function, which takes the output of the physical constraint optimization model as input, and outputs a scalar value for evaluating the merits of the current process conditions by performing a preset mathematical combination operation on the input.

[0039] Fix all network weight parameters of the physical constraint optimization model, and set a set of initial working condition feature vectors and a set of fixed raw material feature vectors;

[0040] The gradient of the process optimization objective function with respect to each operable variable in the operating condition feature vector is calculated using the backpropagation algorithm. This gradient is then used to iteratively update the initial operating condition feature vector until the value of the process optimization objective function no longer improves after several consecutive iterations. At this point, the optimal value is determined, and the operating condition feature vector at this point is used as the optimization operation parameter. The operable variables are the process parameters that constitute the operating condition feature vector and can be directly set in the actual production process.

[0041] A biomass gasification reaction modeling system based on physical information neural networks is used to implement the above method. The system includes:

[0042] Data feature enhancement module: used to acquire raw gasification data characterizing the biomass gasification reaction process, perform data tensor quantization on the raw gasification data, output feature tensors for data enhancement processing, perform adversarial game training on the feature tensors to obtain an adversarial network model, sample the working condition area not covered by the experiment and call the adversarial network model to obtain an enhanced gasification feature dataset.

[0043] Hybrid model pre-construction module: used to perform thermodynamic equilibrium flow subnet calculation based on enhanced gasification feature dataset, obtain theoretical equilibrium prediction result set, construct and train non-equilibrium dynamic flow subnet through theoretical equilibrium prediction result set, call non-equilibrium dynamic flow subnet for calculation, obtain prediction dynamic deviation vector, perform fusion on prediction dynamic deviation vector, obtain final syngas component prediction result set to complete the pre-construction of residual two-flow network architecture;

[0044] Process optimization module: Used to construct a physical constraint loss function to fine-tune the residual two-stream network architecture, obtain a physical constraint optimization model, perform inverse optimization on the physical constraint optimization model, and output optimized operating parameters to guide the gasification process.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] This invention addresses the challenges of insufficient training samples and poor generalization ability in traditional modeling methods due to the scarcity of real experimental data and the systematic bias in simulation data. It utilizes an adversarial network model to perform cross-domain data augmentation on feature tensors derived from experimental and simulation log files. The constructed residual dual-flow network architecture provides a theoretical benchmark through a thermodynamic equilibrium flow subnet and corrects biases through a non-equilibrium dynamic flow subnet, decomposing complex nonlinear prediction into a "benchmark + correction" problem. This overcomes the dual limitations of traditional mechanistic models failing to capture non-equilibrium states and pure data models lacking physical constraints, significantly improving prediction accuracy. The constructed physical constraint loss function fine-tunes the residual dual-flow network architecture using conservation laws and dynamic monotonicity as hard constraints. The resulting physical constraint optimization model fundamentally ensures the physical consistency of prediction results, avoiding the absurd results of "black box" models that violate physical laws. Finally, by performing reverse optimization on the physical constraint optimization model, optimized operating parameters for production guidance are output, achieving a functional closed loop from passive prediction to active process optimization. This provides core technical means for the intelligent and optimized operation of biomass gasification processes. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart of a biomass gasification reaction modeling method based on a physical information neural network provided in an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the blank data structure of the experimental log file and the simulation log file provided in the embodiments of the present invention;

[0050] Figure 3 This invention provides a schematic diagram of the network weight parameter optimization trajectory under a multidimensional physical constraint loss manifold.

[0051] Figure 4 A functional block diagram of a biomass gasification reaction modeling system based on physical information neural networks provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] Please see Figure 1 As shown, this embodiment provides a biomass gasification reaction modeling method based on physical information neural networks, including:

[0055] Step S10: Obtain raw gasification data characterizing the biomass gasification reaction process; perform data tensor quantization on the raw gasification data; output feature tensors for data augmentation processing; perform adversarial game training on the feature tensors to obtain an adversarial network model; sample the operating conditions not covered by the experiment and call the adversarial network model to obtain an enhanced gasification feature dataset.

[0056] Further, step S10 includes:

[0057] Step S11: Obtain raw gasification data characterizing the biomass gasification reaction process, perform data tensor quantization on the raw gasification data, and output feature tensors for data augmentation processing.

[0058] In biomass gasification reaction kinetic modeling, the foundation of the modeling is acquiring and processing raw gasification data describing the entire gasification process. This raw gasification data is a record of the original information for each specific gasification process, existing in a multi-source heterogeneous form. Specifically, the raw gasification data is recorded in two different types of structured text files: experimental log files and simulation log files. The raw gasification data itself consists of two different data categories: static physicochemical properties characterizing the inherent composition of the biomass feedstock before the reaction, and dynamic process control data characterizing the operating state of the gasifier during the reaction. Because the static physicochemical properties and the dynamic process control data differ significantly in physical dimensions, numerical ranges, and mechanisms of influence on the target products, directly inputting unprocessed raw gasification data into a neural network model will lead to difficulty in model convergence and ineffective feature extraction. Therefore, the purpose of constructing a heterogeneous feature mapping space that can uniformly map the heterogeneous raw gasification data into a homogeneous mathematical expression is to transform the discrete, unstructured experimental observation records in the physical world into a standardized numerical matrix that can be efficiently computed by computer processing units, thereby providing a normalized data foundation for subsequent nonlinear feature extraction based on deep learning.

[0059] Specifically, experimental log files and simulation log files are read from the local data acquisition server of the biomass gasification pilot plant or the network file server of the remote simulation computing cluster. The biomass gasification pilot plant is an intermediate-scale gasification reaction system with a complete process flow but scaled-down processing capacity, positioned between small-scale laboratory research and large-scale industrial production. Both the experimental log files and simulation log files are structured text files stored in comma-separated value format. Each line represents an independent record of raw gasification data, and each column corresponds to a specific physical or chemical parameter field. The experimental log files and simulation log files include a set of identical column field names, specifically: moisture mass percentage, ash mass percentage, volatile matter mass percentage, fixed carbon mass percentage, carbon element mass percentage, hydrogen element mass percentage, oxygen element mass percentage, nitrogen element mass percentage, sulfur element mass percentage, chlorine element mass percentage, lower heating value, higher heating value, gasification temperature, gasification pressure, air equivalence ratio, and water vapor to biomass mass ratio. Figure 2 This is a schematic diagram of a blank data structure for an experimental log file and a simulation log file provided in an embodiment of the present invention; for example... Figure 2 As shown in the figure, the internal data structure of the experimental log file and the simulation log file is displayed in tabular form. Each row of the table represents an independent record of raw gasification data, and the first column, "File Type," is used to distinguish the data source. Each column corresponds to a specific physical or chemical parameter field, which is to be filled in in the figure.

[0060] After acquiring the raw gasification data, a heterogeneous feature mapping space and data tensor quantization operations are performed on the raw gasification data. Specifically, a high-dimensional mathematical space, namely the heterogeneous feature mapping space, is defined, which consists of a feedstock attribute subspace and an operating condition parameter subspace. The feedstock attribute subspace is used to quantify the static physicochemical properties in the raw gasification data. Fields characterizing the pyrolysis behavior, mass balance baseline, total chemical energy, and potential corrosion risk of the feedstock are selected from the raw gasification data to form the feedstock attribute subspace. These fields specifically include: moisture mass percentage, ash mass percentage, volatile matter mass percentage, fixed carbon mass percentage, carbon element mass percentage, hydrogen element mass percentage, oxygen element mass percentage, nitrogen element mass percentage, sulfur element mass percentage, chlorine element mass percentage, lower heating value, and higher heating value. Among them, the moisture mass percentage represents the percentage of the total mass of free water and bound water contained in the biomass feedstock during analysis, relative to the total mass of the biomass feedstock. The moisture mass percentage is chosen because water requires a large amount of heat to evaporate in the initial stage of gasification, and its value affects the energy balance and effective reaction temperature of the gasifier. The ash mass percentage is the ratio of the mass of inorganic residue remaining after the biomass feedstock is completely burned in air at a specified temperature (e.g., 815℃) to the total mass of the biomass feedstock. It is chosen because the ash content and its melting characteristics are the main basis for assessing the risk of high-temperature slagging in the gasifier and selecting the slag removal method. The volatile matter mass percentage is the ratio of the total mass of gaseous and liquid products precipitated when the biomass feedstock is heated to a specified temperature (e.g., 900℃) under dry, ash-free conditions to the mass of the biomass feedstock. It is chosen because the volatile matter content is a key indicator characterizing the reactivity of the feedstock during the pyrolysis stage; high volatile matter usually means a faster reaction rate. The fixed carbon mass percentage is calculated under the same baseline, from... The percentage remaining after subtracting moisture, ash, and volatile matter from 100% is chosen because fixed carbon represents the potential of coke gasification, and its content is related to the gas production of subsequent gas-solid reactions. The five fields of carbon mass percentage, hydrogen mass percentage, oxygen mass percentage, nitrogen mass percentage, and sulfur mass percentage represent the mass percentage of each element in the biomass feedstock. They are chosen to provide an atomic-level conservation benchmark for subsequent mass balance calculations. The chlorine mass percentage represents the mass percentage of chlorine in the biomass feedstock. It is chosen because the element is easily converted into corrosive gases at high temperatures and reacts with alkali metals to form gaseous chlorides, which is a key input characteristic for quantifying equipment corrosion risk and downstream catalyst poisoning risk. The lower heating value and higher heating value together characterize the chemical energy contained in the biomass feedstock and are the basis for energy balance calculations. The higher heating value is the total heat released when a unit mass of fuel is completely burned, including the heat of condensation of water vapor in the products, while the lower heating value does not include the heat of condensation of water vapor.The operating parameter subspace consists of independent operational variable fields characterizing reaction temperature, gasification pressure, and the concentration of key reactants. Specifically, it includes gasification temperature, gasification pressure, air equivalence ratio, and the water vapor to biomass mass ratio. Gasification temperature is the arithmetic mean of temperatures at multiple measurement points within the gasifier; it is chosen because temperature is the most significant factor affecting the chemical reaction rate. Gasification pressure is the static pressure measured by a pressure sensor in the free space region at the top of the gasifier; it is chosen because pressure not only affects the density and residence time of gaseous components but also determines the equilibrium position of the reaction where the number of gas moles changes. The air equivalence ratio is the ratio of the actual molar flow rate of air supplied to the gasifier to the molar flow rate of air required for the theoretical stoichiometric complete combustion of the same mass of biomass feedstock; it is a dimensionless number. The air equivalence ratio is chosen because it directly regulates the redox atmosphere in the reaction zone, determining whether the gasification process is exothermic or endothermic. The water vapor to biomass mass ratio is the ratio of the mass flow rate of water vapor supplied to the gasifier per unit time to the mass flow rate of the supplied biomass feedstock. The reason for selecting steam is that steam is a key reactant in the water-gas shift reaction, which determines the ratio of hydrogen to carbon monoxide in the final syngas.

[0061] For each line of raw gasification data parsed from the experimental and simulation log files, normalization is performed to linearly map all values ​​to the closed interval [0,1], eliminating the magnitude differences caused by different physical dimensions. The normalized values ​​are then arranged based on feature sensitivity to determine their order. The aim is to arrange physically correlated features adjacently in the vector space to facilitate the effective extraction of local features, ultimately resulting in a series of standardized feature tensors V. Each feature tensor consists of two concatenated parts: the first part is the raw material feature vector V1, composed of field data constituting the raw material attribute subspace; the second part is the operating condition feature vector V2, composed of field data constituting the operating condition parameter subspace, i.e., V = V1, V2. Each feature tensor is a complete digital description of a specific gasification experiment or simulation, establishing a bijective relationship between physical entity attributes and digital feature space.

[0062] Step S12: Construct a conditional generator and a convolutional discriminator based on the feature tensor, and perform adversarial game training on the conditional generator and the convolutional discriminator to obtain the adversarial network model.

[0063] After outputting the feature tensors, to address the issues of sparse feature tensors from experimental log files failing to fully cover the entire heterogeneous feature mapping space, and large feature tensors from simulation log files exhibiting systematic distributional deviations from the real physical processes, an adversarial network model for cross-domain enhancement of gasification data is constructed through innovative application and combination of traditional neural network models. This adversarial network model is a deep learning architecture comprising two competing sub-networks: a conditional generator and a convolutional discriminator. Its aim is to learn a nonlinear mapping function from the distribution of simulation data to the distribution of real experimental data, thereby generating a large number of high-quality virtual data samples with statistical properties consistent with real experimental data to fill the data gaps in the high-dimensional feature space of the experimental source feature tensors. Its scientific basis lies in the fact that while simulation models can accurately capture the thermodynamic equilibrium trend under ideal conditions, they cannot simulate non-ideal, non-equilibrium kinetic effects such as catalyst micro-poisoning caused by impurities like the percentage of chlorine by mass, or localized heat transfer deterioration within the furnace caused by the percentage of ash by mass. The adversarial network model aims to implicitly learn non-ideal correction functions through a data-driven approach.

[0064] Specifically, a condition generator and a convolutional discriminator are constructed. The network structure of the condition generator is a deep residual network, whose input consists of two parts: a simulation source condition vector and a high-dimensional noise vector. This deep residual network consists of four cascaded residual blocks, each containing two fully connected layers with 256 neurons per layer. The activation function is ReLU. The simulation source condition vector is obtained by concatenating all field data constituting the raw material attribute subspace (i.e., the raw material feature vector) and all field data constituting the operating condition parameter subspace (i.e., the operating condition feature vector) from the feature tensor of the simulation log file. In other words, the simulation source condition vector includes all input features from the feature tensor of the simulation log file. The purpose is to provide a reasonable starting point for the operating conditions. For example, a typical simulation source condition vector is selected as the input. For instance, the specific operating condition values ​​contained in the simulation source condition vector are: gasification temperature of 850 degrees Celsius, air equivalence ratio of 0.28, and moisture mass percentage of 15%. The high-dimensional noise vector is randomly sampled from a preset standard normal distribution with a mean of 0 and a variance of 1 by a pseudo-random number generation module within the computing device used in biomass gasification reaction kinetic modeling. Its dimension is set to be the same as the dimension of the simulation source condition vector; for example, the noise vector's dimension is set to 100. This injects randomness, allowing the condition generator to generate multiple virtual samples with slight differences but equally reasonable results even for the same simulation source condition vector. This simulates the "data scattering under the same operating conditions" phenomenon caused by uncontrollable factors such as measurement errors and minor environmental fluctuations in real experiments. After receiving the simulation source condition vector and the high-dimensional noise vector, the deep residual network performs a series of nonlinear transformations within the deep residual network, ultimately outputting a model... A virtual high-fidelity feature tensor with the same data structure and dimensions as the simulated source condition vector is used. The convolutional discriminator determines whether the input data originates from a real data distribution or a virtual data distribution generated by the condition generator. The network structure uses a multi-layer convolutional neural network with LeakyReLU (Leaky Corrected Linear Unit) as the internal activation function. Its input is a feature tensor from any data source. After processing by the multi-layer convolutional neural network, its output is fed into a final fully connected layer. The final fully connected layer has only one output neuron and does not apply any activation function. The final output is a scalar value, which reflects the convolutional discriminator's evaluation of the "realism" of the input feature tensor. The larger the value, the closer it is to the real distribution.

[0065] After constructing the conditional generator and convolutional discriminator, adversarial game training is performed. Specifically, all feature tensors are divided into feature tensors whose data source is experimental log files and simulation log files, defined as simulation source feature tensors and experimental source feature tensors, respectively. Before adversarial game training, all experimental source feature tensors whose data source is experimental log files are divided into training subsets and validation subsets according to a ratio, for example, 80% and 20%; all simulation source feature tensors whose data source is simulation log files are defined as the conditional input set. The training subset is used by the convolutional discriminator to learn real samples of the actual data distribution. The conditional input set serves as a reference blueprint for the conditional generator to generate virtual samples. The validation subset does not participate in training at all; its role is to periodically evaluate the performance of the conditional generator during training. N feature tensors are randomly selected from the simulated source feature tensor and the experimental source feature tensor, respectively named the simulated source feature tensor batch and the experimental source feature tensor batch. The value of N is determined based on the hardware resources of the computing device, such as the video memory capacity of the graphics processor. For example, N is set to 256. Simultaneously, the computing device generates a batch of high-dimensional noise vectors comprising N high-dimensional noise vectors. Each simulated source feature tensor in the simulated source feature tensor batch is concatenated with the corresponding high-dimensional noise vector in the high-dimensional noise vector batch along the feature dimension to form a combined input tensor. This combined input tensor is used as input to the conditional generator. After processing by the deep residual network of the conditional generator, a virtual high-fidelity feature tensor batch is output. In the initial training phase, such as the 10th iteration, the conditional generator has not yet converged, and the output virtual data deviates significantly from physical reality. For example, the generated volume fractions of the main components are: 12% carbon monoxide, 8% hydrogen, 25% carbon dioxide, and 1% methane. These values ​​are far lower than normal gasification levels and do not conform to material balance. The virtual high-fidelity feature tensor batch has the same dimensions and data structure as the simulated source feature tensor batch. The virtual high-fidelity feature tensor batch and the experimental source feature tensor batch are used as two independent inputs to the convolutional discriminator. The convolutional discriminator processes the data and outputs a scalar value for the feature tensors of the virtual high-fidelity feature tensor batch and the experimental source feature tensor batch, respectively, thus forming two scoring batches: the generated data scoring batch corresponding to the virtual high-fidelity feature tensor batch and the real data scoring batch corresponding to the experimental source feature tensor batch.

[0066] After obtaining two scoring batches, convolutional discriminator optimization and conditional generator optimization are performed. The convolutional discriminator optimization treats a set of trainable network weight parameters—used to define the nonlinear transformation capability of the deep residual network within the conditional generator—as constants. The convolutional discriminator loss function L1 is calculated, which consists of the algebraic sum of three components: the arithmetic mean A of the data scoring batches, the inverse of the arithmetic mean -B of the real data scoring batches, and a gradient penalty term C. A and -B are obtained using the arithmetic mean formula, and the gradient penalty term is set according to Lipschitz constraints to ensure that the convolutional discriminator approximates 1 on the data manifold, thereby stabilizing the training process. Specifically, between the real data batch and the generated data batch, element-wise linear interpolation is performed using a random number uniformly distributed in the interval [0,1] to generate a batch of interpolated samples. The square of each component of the gradient is summed, and then the square root is taken to obtain the norm of the gradient. The square of the difference between this norm and 1 is calculated to obtain the gradient penalty term. Finally, the expression for the convolutional discriminator loss function is: L1 = A − B + λ × C, where λ is a preset penalty coefficient, used to balance the weight of the gradient penalty term in the convolutional discriminator loss function. The value is set based on achieving a balance between the stability and convergence speed of adversarial training. For example, the penalty coefficient is set to 10. After calculating the specific value of the convolutional discriminator loss function, the partial derivatives of the network weight parameters inside the convolutional discriminator are calculated using the backpropagation algorithm. The optimizer updates the weights along the negative direction of the gradient, with the goal of minimizing the convolutional discriminator loss function, thereby maximizing the difference between the real score and the generated score in terms of effect. The execution process of the conditional generator optimization is as follows: the trainable network weight parameters contained in the multi-layer convolutional neural network inside the convolutional discriminator are treated as constants, and the conditional generator loss function is calculated. The conditional generator loss function is the negative of the arithmetic mean of the data score batch, i.e., -A. After obtaining the specific value of the conditional generator loss function, the partial derivatives with respect to the weight parameters of the internal network of the conditional generator are calculated using the backpropagation algorithm. The weights are then updated along the negative direction of the gradient using the optimizer. The goal is to minimize the conditional generator loss function, thereby maximizing the generated data score batches that can be obtained by the virtual high-fidelity feature tensor in terms of effect.

[0067] The convolutional discriminator optimization and conditional generator optimization are performed alternately according to a preset alternation frequency. This alternation frequency is set to ensure that the learning ability of the convolutional discriminator can rapidly improve in the early stages of training, providing accurate and effective gradient information to the conditional generator, while maintaining a dynamic balance between the conditional generator and the convolutional discriminator in the later stages of training. For example, the alternation frequency is set to 5, meaning that after every 5 convolutional discriminator optimizations, a conditional generator optimization is performed once; this addresses the issue of the convolutional discriminator easily converging prematurely.

[0068] Taking a training batch size of 256, a gradient penalty coefficient of 10, and a total training iteration count of 5000 rounds as an example, the specific training process is as follows: All feature tensors are divided into simulated source feature tensors and experimental source feature tensors based on their data source. The experimental source feature tensors are further divided into training subsets and validation subsets, with 80% and 20% respectively. The simulated source feature tensors are defined as the conditional input set. In each training iteration cycle, 256 feature tensors are randomly selected from both the simulated and experimental source feature tensors to form simulated source feature tensor batches and experimental source feature tensor batches. Simultaneously, a high-dimensional noise vector batch containing 256 high-dimensional noise vectors is generated. The simulated source feature tensor batch and the high-dimensional noise vector batch are concatenated and input into the conditional generator to output a virtual high-fidelity feature tensor batch. Alternating optimization is performed on the convolutional discriminator loss function, with one conditional generator optimization performed after every five convolutional discriminator optimizations.

[0069] In the early stages of training, from round 1 to round 500, the conditional generator had not yet converged, and the output virtual data deviated significantly from physical reality. At this time, the convolutional discriminator could easily distinguish between real and fake data, and the loss value of the convolutional discriminator's loss function rapidly decreased and remained at a low level, while the loss value of the conditional generator's loss function was high and fluctuated wildly. For example, in the 10th iteration, although the generated virtual high-fidelity feature tensor contained normalized raw material and operating condition features in its data structure, its corresponding physical meaning was severely distorted: for example, the volume fraction of carbon monoxide was 12%, the volume fraction of hydrogen gas was 8%, the volume fraction of carbon dioxide was 25%, and the volume fraction of methane was 1%, which were far lower than the normal gasification level and did not conform to material balance. In the middle of training, from round 500 to round 3000, as the adversarial game training deepened, the conditional generator began to capture the statistical characteristics of real data. As the difficulty of distinguishing between true and false data increases, the loss value of the convolutional discriminator's loss function begins to show an oscillating upward trend, while the loss value of the conditional generator's loss function gradually decreases. The data distribution of the generated data begins to partially overlap with the data distribution of the real experimental data. Towards the end of training from the 3000th to the 5000th round, the system enters the Nash equilibrium search phase, and its performance stabilizes until the total loss function converges below the preset adversarial threshold, at which point significant unidirectional drift no longer occurs. At this point, for the same simulated source conditions, the virtual high-fidelity feature tensor output by the conditional generator possesses a high degree of physical realism. This feature tensor specifically includes normalized encoded raw material properties such as the content of carbon, hydrogen, and oxygen elements, and operating parameters. For example, its corresponding component prediction values ​​are: carbon monoxide volume fraction 23%, hydrogen gas volume fraction 17%, carbon dioxide volume fraction 15%, and methane volume fraction 3%. This conforms to thermodynamic laws and element conservation, and accurately reflects the component deviations caused by non-ideal factors (such as localized heat transfer deterioration due to ash content) in real experiments. Training terminates when the training reaches a preset 5000 iterations or the total loss function meets the convergence condition, outputting a trained adversarial network model with cross-domain mapping capabilities. The total loss function is a weighted sum of the convolutional discriminator loss function and the condition generator loss function, used to monitor the overall convergence state of the entire adversarial game training process. The adversarial threshold serves as a standard for determining whether the adversarial game training process has reached a stable state and can be terminated early; it is a small value set before training on the validation subset that represents a stable convergence stage.

[0070] Step S13: Sample the working area not covered by the experiment and call the adversarial network model to obtain the enhanced gasification feature dataset.

[0071] After outputting the trained adversarial network model, in order to finally solve the problem of sparse feature tensors from the experimental log files, the condition generator inside the adversarial network model is used to generate high-quality virtual data on a large scale and at low cost, constructing an enhanced gasification feature dataset with broad coverage and realistic distribution, providing sufficient data support for subsequent work.

[0072] Sampling is performed on areas not covered by the experiment. Specifically, in the heterogeneous feature mapping space, operating condition areas not covered by experimental data are identified. This involves visualizing all feature tensors from the experimental log files in the heterogeneous feature mapping space to determine the boundaries of the experimental data point distribution. For example, a scatter plot is drawn after dimensionality reduction using principal component analysis. Areas outside these boundaries but within a reasonable physical operating range, such as gasification temperatures between 800℃ and 1200℃ (but the experiment only covers up to 1000℃), are sampled between 1000℃ and 1200℃. A series of new operating condition feature vectors are generated through grid sampling, defined as target operating condition feature vectors. Simultaneously, the raw material feature vector portion is extracted from all experimental source feature tensors originating from the experimental log files. The arithmetic mean of each dimension of the raw material feature vector is calculated using the arithmetic mean formula to obtain an average raw material feature vector. The target operating condition feature vector is concatenated with the average raw material feature vector to form a batch of simulation source condition vectors for areas not covered by the experiment, defined as uncovered simulation vectors. The reason for using the average raw material feature vector as the generation condition is that it represents the most commonly used and representative raw material types in real experiments, and has stronger guiding significance for engineering practice.

[0073] After obtaining the uncovered simulation vectors, a batch of new high-dimensional noise vectors is generated using computing equipment. Following the method for generating combined input tensors, the uncovered simulation vectors are concatenated with the new high-dimensional noise vectors to obtain a new combined input tensor, defined as the uncovered input tensor. The conditional generator trained in the adversarial network model is then invoked, using the uncovered input tensor as its input. The conditional generator outputs a batch of virtual high-fidelity feature tensors for data augmentation, defined as augmented feature tensors. These augmented feature tensors are merged with all feature tensors to obtain an enhanced gasification feature dataset. This dataset includes both experimental source feature tensors from the experimental log files and simulation source feature tensors from the simulation log files, as well as augmented feature tensors from regions not covered by the experiments. This solves the core problems of insufficient training data and incomplete coverage of operating conditions caused by high experimental costs and long cycles.

[0074] Step S10 addresses the technical challenges of insufficient training samples and incomplete operational coverage in traditional modeling methods due to the sparse data volume of experimental log files and the systematic distribution bias in simulation log files. This is achieved by using raw gasification data, data tensorization operations, and an adversarial network model. This enables the construction of an enhanced gasification feature dataset covering a wider range of operational conditions for subsequent high-fidelity modeling. Specifically, data tensorization unifies the heterogeneous raw gasification data from multiple sources into standardized feature tensors; adversarial game training learns the nonlinear mapping from simulation data distribution to real experimental data distribution, resulting in an adversarial network model; and sampling and invoking the adversarial network model for areas not covered by the experiment specifically generates enhanced feature tensors, filling the gaps in the high-dimensional operational space of the experimental data.

[0075] Step S20: Perform thermodynamic equilibrium flow subnet calculation based on the enhanced gasification feature dataset to obtain the theoretical equilibrium prediction result set. Construct and train the non-equilibrium dynamic flow subnet using the theoretical equilibrium prediction result set. Call the non-equilibrium dynamic flow subnet for calculation to obtain the predicted dynamic deviation vector. Perform fusion on the predicted dynamic deviation vector to obtain the final syngas component prediction result set and complete the pre-construction of the residual two-flow network architecture.

[0076] Further, step S20 includes:

[0077] Step S21: Perform thermodynamic equilibrium flow subnet calculations based on the enhanced gasification feature dataset to obtain a theoretical equilibrium prediction result set.

[0078] After outputting the enhanced gasification feature dataset, to address the problem that traditional single neural network models are prone to getting trapped in local optima or producing predictions that violate physical laws due to the lack of prior physical knowledge constraints when predicting complex physicochemical processes, a residual dual-flow network architecture is constructed. This residual dual-flow network architecture comprises two parallel data processing paths: a thermodynamic equilibrium flow subnet and a non-equilibrium kinetic flow subnet. The core of this step is to construct and execute the thermodynamic equilibrium flow subnet, aiming to calculate a physically reasonable theoretical benchmark value based on first principles of thermodynamics for each operating condition in the enhanced gasification feature dataset, providing a stable anchor point for the subsequent non-equilibrium kinetic flow subnet.

[0079] Specifically, the thermodynamic equilibrium flow subnet is logically a computational module that does not contain any trainable neural network weight parameters. Instead, it embeds a numerical solution program for the Gibbs free energy minimization algorithm. The scientific basis of this algorithm is a corollary of the second law of thermodynamics, which states that in a closed, isothermal, and isobaric system, chemical reactions will spontaneously proceed in the direction of decreasing Gibbs free energy until a minimum is reached, at which point the system is in chemical equilibrium. The processing procedure of the thermodynamic equilibrium flow subnet is as follows: each feature tensor is read one by one from the enhanced gasification feature dataset. For each read feature tensor, the operating condition feature vector and the feed feature vector are extracted to construct a closed reaction system for subsequent thermodynamic calculations. Specifically, the values ​​of gasification temperature and gasification pressure are extracted from the operating condition feature vector. The reason for this selection is that the Gibbs free energy minimization algorithm is a function of temperature and pressure, which are the two most fundamental thermodynamic state parameters for the chemical reaction equilibrium state. The values ​​of air equivalence ratio and water vapor to biomass mass ratio are also extracted. The mass percentages of carbon, hydrogen, oxygen, nitrogen, and sulfur are extracted from the raw material feature vector. These five fields are defined as elemental analysis data. The reason for this selection is that the solution process of the Gibbs free energy minimization algorithm strictly adheres to the law of conservation of atoms, providing a basis for calculating the initial total number of atoms in the reaction system.

[0080] The number of moles for each field is calculated based on elemental analysis data. For example, the number of moles for the carbon mass percentage is obtained by dividing the carbon mass percentage by the standard atomic weight of carbon. Based on the values ​​of the air equivalence ratio and the water vapor to biomass mass ratio in the operating condition feature vector, the number of extra atomic moles introduced into the reaction system by air and water vapor is calculated using stoichiometric formulas. By adding the number of atoms of the raw materials themselves to the number of atoms introduced by the gasifying agent, the total atomic conservation constraint of the entire reaction system in the initial state can be obtained. Using the total atomic conservation constraint, vaporization temperature, and vaporization pressure as input parameters, the algorithm for minimizing Gibbs free energy within the thermodynamic equilibrium flow network is used for calculation. Specifically, based on the atomic composition of the input parameters, the possible gaseous products generated in the reaction system are determined. These gaseous products are defined as chemical substances that exist stably in gaseous form at a given vaporization temperature and pressure. A thermodynamic database containing the thermodynamic properties of each product is invoked to find the solution that minimizes the total Gibbs free energy of the system. This minimum solution is a set of values ​​describing the molar number of each gaseous product at equilibrium. These molar numbers are converted into their respective volume fractions in the total gas, resulting in a theoretical equilibrium prediction vector. This vector includes the predicted volume fractions of hydrogen, carbon monoxide, carbon dioxide, and methane generated under ideal equilibrium conditions. The total gas refers to the mixture of all products existing in gaseous form in the reaction system at thermodynamic equilibrium. The theoretical equilibrium prediction vectors corresponding to each feature tensor in the enhanced vaporization feature dataset are combined to obtain a theoretical equilibrium prediction result set. For example, consider a specific feature tensor as an input parameter: assuming the extracted vaporization temperature is 900 degrees Celsius, the vaporization pressure is 0.1 MPa, and the raw material baseline mass is 1 kg, the elemental analysis data shows that the raw material itself contains 41.6 moles of carbon, 60.0 moles of hydrogen, and 26.5 moles of oxygen. Simultaneously, with an air equivalence ratio of 0.3 and a water vapor ratio, the additional 15.2 moles of oxygen and 58.8 moles of nitrogen introduced by the vaporizing agent are calculated. Adding the raw material's own atoms to the atoms introduced by the vaporizing agent yields the total atomic conservation constraint vector for the entire reaction system, i.e., the total molar amount of carbon, hydrogen, oxygen, nitrogen, and sulfur atoms. This total atomic conservation constraint vector, along with the vaporization temperature and pressure, is input into the Gibbs free energy minimization algorithm. Specifically, a total Gibbs free energy function is constructed, defined as the sum of the products of the chemical potentials of hydrogen, carbon monoxide, carbon dioxide, methane, water vapor, and nitrogen in the reaction system and their corresponding molar numbers.Based on the law of conservation of mass, a linear equation constraint is established for each chemical element. Taking carbon as an example, the equation logic is: the number of moles of carbon monoxide multiplied by one, plus the number of moles of carbon dioxide multiplied by one, plus the number of moles of methane multiplied by one equals the total number of carbon atoms in the previously calculated total atomic conservation constraint vector, which is 41.6 moles. Similarly, linear equation constraints are established for hydrogen, oxygen, and nitrogen. The number multiplied by the number of moles of each component strictly corresponds to the number of atoms of that specific element in the chemical formula of that component. Specifically, for hydrogen, since the molecular structures of hydrogen gas, methane, and water vapor contain 2, 4, and 2 hydrogen atoms respectively, the equation logic for the linear equation constraint of hydrogen is: the number of moles of hydrogen gas multiplied by two, plus the number of moles of methane multiplied by four, plus the number of moles of water vapor multiplied by two, and the algebraic sum equals... The total number of moles of hydrogen atoms; for oxygen, since the molecules of carbon monoxide, carbon dioxide, and water vapor contain 1, 2, and 1 oxygen atom respectively, the logical equation is: the number of moles of carbon monoxide multiplied by 1, plus the number of moles of carbon dioxide multiplied by 2, plus the number of moles of water vapor multiplied by 1, the algebraic sum equals the total number of moles of oxygen atoms; for nitrogen, since the molecular structure of nitrogen gas contains 2 nitrogen atoms, the logical equation is: the number of moles of nitrogen gas multiplied by 2, the value equals the total number of moles of nitrogen atoms. All the above equations together form a system of constraint equations. The Gibbs free energy minimization algorithm introduces the Lagrange multiplier method, combining the total Gibbs free energy function with the constraint equations. An unconstrained Lagrange function is formed by the chemical potential term of each component and the multiplier terms corresponding to the constraint terms of each element. The partial derivatives of each variable in the Lagrange function are calculated and set to zero, thus obtaining a set of nonlinear algebraic equations. The Newton-Raphson iteration method is used to numerically solve the nonlinear equations. When the difference between the total Gibbs free energy of the system calculated in two adjacent iterations is less than the preset convergence accuracy, such as 10, the algorithm is considered successful. -6 When convergence is achieved, the solution to the equation system is the molar number of each component in thermodynamic equilibrium. After iterative convergence, the molar number of each component in equilibrium is output, for example, 32.5 moles of hydrogen, 28.4 moles of carbon monoxide, 9.6 moles of carbon dioxide, 7.8 moles of water vapor, 29.4 moles of nitrogen, and 0.01 moles of methane. Dividing the molar number by the total number of moles of the gaseous products converts it to a normalized volume fraction, yielding the theoretical equilibrium prediction vector corresponding to this characteristic tensor.

[0081] Step S22: Construct and train the non-equilibrium dynamic subnet based on the theoretical equilibrium prediction result set.

[0082] After outputting the theoretical equilibrium prediction result set, to address the issue of systematic deviations between the theoretical equilibrium prediction vector and the actual physical process due to neglecting reaction kinetic constraints such as insufficient residence time and uneven gas-solid mixing, a non-equilibrium kinetic flow subnet is constructed and trained. Its function is to learn and quantify these deviations caused by non-ideal, non-equilibrium effects, thereby accurately correcting the theoretical equilibrium prediction vector.

[0083] The non-equilibrium dynamics subnet is a specific computational entity. Its internal network structure uses a deep residual network, whose input layer is designed to receive a joint input vector. This joint input vector is composed of two concatenated data parts: the first part is each feature tensor in the enhanced gasification feature dataset itself; the second part is the theoretical equilibrium prediction vector corresponding to the feature tensors in the enhanced gasification feature dataset within the theoretical equilibrium prediction result set. After receiving the joint input vector, the deep residual network processes it through multiple cascaded deep residual blocks before sending it to a final fully connected layer. This fully connected layer maps a dynamic bias vector with the same dimension as the theoretical equilibrium prediction vector; that is, the dynamic bias vector is a prediction correction vector for the theoretical equilibrium prediction vector. The training operation is performed specifically to provide supervision signals for the training of the non-equilibrium dynamics subnet. A target deviation dataset is constructed. This dataset is generated by extracting a target state vector from each feature tensor in the enhanced gasification feature dataset. The target state vector is the numerical value of the four fields: hydrogen gas fraction, carbon monoxide volume fraction, carbon dioxide volume fraction, and methane volume fraction, of the theoretical equilibrium prediction vector corresponding to the feature tensor. The target state vector is then subtracted element-wise from the corresponding theoretical equilibrium prediction vector; the resulting difference vector is the target deviation vector. Combining the target deviation vectors calculated from all feature tensors yields the target deviation dataset. Before training, the enhanced gasification feature dataset and the corresponding target deviation dataset are randomly divided into a training set and a validation set according to a preset ratio, such as 80%:20%. The training set is used to update the network weight parameters during training; the validation set does not participate in weight updates but is only used to periodically evaluate generalization ability and monitor for overfitting. A loop containing multiple training iterations is executed. In each training iteration, K joint input vectors and K target bias vectors corresponding to the joint input vectors are randomly selected from the training set to form a training batch. The value of K is determined based on the GPU memory capacity of the computing device, ensuring that the data for a single training iteration can be fully loaded into the GPU memory without overflow. A value as large as possible is chosen to obtain more stable gradient estimation; for example, K is set to 512. The K joint input vectors from the training batch are used as input to a non-equilibrium dynamics subnet. After processing by a deep residual network within the non-equilibrium dynamics subnet, a prediction bias batch containing K prediction dynamics bias vectors is output. The mean squared error between the prediction dynamics bias vectors in the prediction bias batch and the corresponding target bias vectors in the training batch is calculated, and this mean squared error is used as the loss function value for the current iteration.The partial derivatives of the loss function with respect to the weight parameters within the non-equilibrium dynamic subnet are calculated using the backpropagation algorithm, and an optimizer is used to update the weights once. To balance the computational cost of training with the timeliness of monitoring for model overfitting, a fixed validation frequency is set to perform validation operations periodically. The validation frequency is defined as the number of training iterations after which model performance is evaluated; for example, the validation frequency is set to 100. After the training iteration cycle with the validation frequency limit is reached, a validation operation is performed. This involves inputting all joint input vectors from the validation set into the non-equilibrium dynamics subnet and calculating the loss function value on the validation set. After each validation operation, the current validation set loss function value is compared with the previous loss function value. If performance saturation or maximum cycle termination occurs, training is terminated, resulting in a trained non-equilibrium dynamics subnet. During the above training iteration process, the loss function value decreases in stages with the number of iterations. For example, taking 5000 as the total number of iterations, in the early stage of training from the 1st to the 500th round, the loss value drops rapidly from its initial high value, reflecting the network's rapid capture of large-scale biases. In the middle stage of training from the 500th to the 2000th round, the loss values ​​of the validation set and the training set remain highly synchronized. In the late stage of training from the 2000th to the 5000th round, the loss value tends to level off, and after some minor local oscillations, it finally converges at a low value. To further disclose the mapping and computation logic during training, for example, in any training iteration, a joint input vector is extracted from the training set. This vector may contain, for example, normalized feature tensors such as a vaporization temperature of 900 degrees Celsius and an air equivalence ratio of 0.3, and a corresponding theoretical equilibrium prediction vector, such as 40.5% hydrogen, 35.4% carbon monoxide, 12.0% carbon dioxide, and 0.01% methane. After inputting the joint input vector into the non-equilibrium kinetic flow subnet, a predicted kinetic bias vector is output, for example, -5.5% hydrogen, -4.2% carbon monoxide, +3.0% carbon dioxide, and +6.5% methane. Simultaneously, a corresponding target bias vector is extracted from the target bias dataset, for example, -5.4% hydrogen, -4.3% carbon monoxide, +3.1% carbon dioxide, and +6.4% methane. The computing device then subtracts the predicted kinetic bias vector from the target bias vector element-wise and sums the squares to obtain the mean squared error contribution value, which is used to drive the backpropagation algorithm to update the network weights.The performance saturation termination refers to the termination of training when the relative rate of change of the validation set loss function value is less than a preset convergence accuracy threshold for M consecutive validation operation cycles. The M value is set to avoid premature training termination due to accidental fluctuations in a single validation cycle; for example, M is set to 10. The convergence accuracy threshold is set based on the desired final model accuracy; for example, the convergence accuracy threshold can be set to 0.01%. The relative rate of change is calculated by dividing the absolute value of the difference between the current validation operation's loss function value and the loss function value of the previous validation operation cycle by the loss function value of the previous validation operation cycle. Maximum cycle termination is used to prevent indefinite computational resource consumption due to improper training parameter settings that prevent training operations from ever converging. A maximum number of training iterations is preset as the upper limit of the training operation; for example, the maximum number of training iterations is set to 5000. When the training iteration cycle reaches this upper limit, training is forcibly terminated regardless of whether convergence has occurred.

[0084] Step S23: The non-equilibrium dynamic flow subnet is called for calculation to obtain the predicted dynamic deviation vector. The predicted dynamic deviation vector is fused to obtain the final syngas component prediction result set, thus completing the pre-construction of the residual two-flow network architecture.

[0085] After obtaining the theoretical equilibrium prediction result set and the trained non-equilibrium dynamic flow subnet, a fusion calculation operation is performed to obtain a final prediction result that incorporates both theoretical constraints and non-equilibrium dynamic corrections. The aim is to combine the theoretical benchmark value provided by the thermodynamic equilibrium flow subnet with the deviation correction amount learned by the non-equilibrium dynamic flow subnet, thereby outputting a high-precision final prediction value.

[0086] Specifically, for each feature tensor in the enhanced gasification feature dataset, the corresponding theoretical equilibrium prediction vector is extracted from the theoretical equilibrium prediction result set and concatenated to obtain a prediction input vector. This prediction input vector is used as the input to the non-equilibrium kinetic flow subnet. After processing by the deep residual network inside the non-equilibrium kinetic flow subnet, a prediction kinetic deviation vector is output. The theoretical equilibrium prediction vector and the prediction kinetic deviation vector are added to obtain the syngas component prediction vector. The syngas component prediction vector includes four elements: the final predicted hydrogen gas integral, carbon monoxide volume fraction, carbon dioxide volume fraction, and methane volume fraction. Each element is obtained by adding the corresponding element in the theoretical equilibrium prediction vector to the corresponding element in the prediction kinetic deviation vector. For example, the final predicted hydrogen gas integral is equal to the element value representing the hydrogen gas integral in the theoretical equilibrium prediction vector plus the element value representing the correction amount of the hydrogen gas integral in the prediction kinetic deviation vector. By repeatedly performing the above operation on each feature tensor in the enhanced gasification feature dataset, a complete final syngas component prediction result set, corresponding one-to-one with each feature tensor in the enhanced gasification feature dataset, is output. This final syngas component prediction result set is the final output of the entire residual two-stream network architecture. The residual two-stream network architecture is a hybrid computational framework integrating a mechanistic model and a data-driven model. It consists of three core components: a thermodynamic equilibrium flow subnet for calculating theoretical equilibrium prediction results; a non-equilibrium kinetic flow subnet for learning non-equilibrium kinetic deviations (the non-equilibrium kinetic deviations are implemented by the non-equilibrium kinetic flow subnet); and the final fusion computation operation in this step, which combines the outputs of the thermodynamic equilibrium flow subnet and the non-equilibrium kinetic flow subnet. This residual two-stream network architecture successfully decomposes a complex nonlinear prediction problem into a linear baseline plus correction problem. This ensures that the final prediction results neither deviate from the fundamental laws of physicochemical science nor fail to accurately capture the dynamic changes caused by various non-ideal factors in real industrial processes, thus achieving higher prediction accuracy and stronger generalization ability than any single model.

[0087] Step S20, by performing thermodynamic equilibrium flow subnet calculations, non-equilibrium dynamic flow subnet calculations, and predicting dynamic deviation vectors, solves the technical problem of inaccurate prediction results and poor generalization ability caused by traditional single models neglecting physical law constraints or failing to capture non-equilibrium dynamic effects, thus realizing the pre-construction of the residual dual-flow network architecture. Specifically, the thermodynamic equilibrium flow subnet calculation provides a set of theoretical equilibrium state prediction results based on first principles; the training of the non-equilibrium dynamic flow subnet focuses on learning the deviation between the theoretical equilibrium state prediction vector and the actual physical process, resulting in a non-equilibrium dynamic flow subnet capable of quantifying non-ideal effects.

[0088] Step S30: Construct a physical constraint loss function to fine-tune and train the residual two-stream network architecture to obtain a physical constraint optimization model. Perform reverse optimization on the physical constraint optimization model and output optimized operating parameters to guide the gasification process.

[0089] Further, step S30 includes:

[0090] Step S31: Construct a physical constraint loss function, and use the physical constraint loss function to fine-tune and train the residual two-stream network architecture to obtain a physical constraint optimization model.

[0091] After obtaining the final syngas component prediction result set, in order to further improve the extrapolation prediction capability of the residual two-stream network architecture when facing operating conditions not covered by the training data, and to ensure that any output strictly follows the fundamental laws of physicochemical processes, known physical laws, such as the conservation of mass and energy, are transformed into mathematically differentiable penalty terms to construct a physical constraint loss function. This physical constraint loss function is then used to perform final fine-tuning training on the already trained residual two-stream network architecture. The aim is to force the model's optimization trajectory to fall on a manifold that conforms to physical laws, thereby endowing it with inherent physical consistency.

[0092] The physical constraint loss function consists of two parts. The first part is the conservation law constraint term, which includes the elemental mass conservation residual and the system enthalpy change residual. The elemental mass conservation residual is calculated as follows: for each feature tensor in the enhanced gasification feature dataset, the initial total elemental mass of the reaction system is calculated. The initial total elemental mass of the reaction system refers to the sum of the masses of all chemical elements contained in all substances entering the reaction system before the reaction begins, i.e., the biomass feedstock and gasifying agent. This total mass is obtained by applying the same stoichiometric relationship as in step S21 to the feedstock feature vector and the operating condition feature vector contained in the feature tensor. The total elemental mass of the predicted products is then calculated. The total elemental mass of the predicted products refers to the sum of the masses of all chemical elements contained in all products predicted by the residual dual-stream network architecture and by calling a tar and semi-coke prediction sub-model after the reaction ends. The reason for using the tar and semi-coke prediction sub-model is that the actual biomass gasification process is not an ideal process that can completely convert all carbon elements into gaseous products. Due to reaction kinetic limitations and non-equilibrium effects, a portion of liquid tar and a portion of solid semi-coke will be generated. The tar is a complex mixture of various organic compounds produced during the gasification process that can be condensed at room temperature, while the semi-coke is a solid residue rich in fixed carbon remaining after the biomass feedstock has undergone pyrolysis and gasification. Since the primary prediction target of the residual two-stream network architecture is the gas phase product components, additional tar and semi-coke prediction sub-models are invoked to estimate the yields of these two products, thereby completing the full component mass balance calculation for all products. The predicted masses of each element in the gas phase, tar, and semi-coke are summed to obtain the total elemental mass of the predicted products. The absolute value of the difference between the initial total elemental mass of the reaction system and the total elemental mass of the predicted products is taken as the elemental mass conservation residual. For example, in a specific training iteration, the initial total elemental mass of the reaction system is calculated based on the raw material feature vector and the operating condition feature vector. Taking carbon as an example, the carbon element is 1200. However, the total elemental mass of the predicted products obtained by the residual two-stream network architecture and the tar and semi-coke prediction sub-models is 1250, which obviously violates the law of mass conservation. At this time, the calculated elemental mass conservation residual is 1200 minus the absolute value of 1250, which is 50.The calculation method for the system enthalpy change residual is as follows: Based on the input reaction temperature and vaporization pressure, the thermodynamic database is called to calculate the total enthalpy of reactants and the total enthalpy of predicted products. The total enthalpy of reactants is the total enthalpy value of all substances entering the reaction system at a given initial temperature before the reaction begins. The vaporization temperature value is extracted from the operating condition feature vector as the initial temperature of reactants entering the reaction system. The total enthalpy of reactants is then queried and calculated from the thermodynamic database. The total enthalpy of predicted products refers to the total enthalpy value of all products predicted by the residual dual-flow network architecture and the tar and semi-coke prediction sub-models at the vaporization temperature after the reaction ends. The composition and molar amount of all predicted products are determined by the syngas component prediction vector and the predicted tar mass yield output by the tar and semi-coke prediction sub-models. The total enthalpy of predicted products is then calculated by querying the standard enthalpy of formation and specific heat capacity data of each product from the thermodynamic database. The calculated total enthalpy of reactants is subtracted from the total enthalpy of predicted products to obtain the net enthalpy change of the reaction. To ensure that the energy balance calculation takes into account actual non-adiabatic effects, the net enthalpy change is subtracted from a wall heat loss parameter. The final difference is the system enthalpy change residual. The wall heat loss parameter is a value characterizing the amount of heat lost to the environment through the reactor wall under current operating conditions. This value is a known input parameter obtained through conventional engineering testing methods, such as heat flow meter measurement or estimation. For example, given a specific operating condition of 850 degrees Celsius vaporization temperature and 0.1 MPa vaporization pressure, the total enthalpy of the reactants calculated from a thermodynamic database is 500 MJ / h, and the known wall heat loss parameter measured by a heat flow meter is 15 MJ / h, assuming the predicted total enthalpy of the products is 480 MJ / h, this means there is a discrepancy between the input energy (i.e., the total enthalpy of the reactants) and the sum of the output energy and losses (i.e., the sum of the predicted total enthalpy of the products and the wall heat loss parameter), violating the law of conservation of energy.

[0093] The second part is the kinetic monotonicity constraint. This constraint is based on Arrhenius's law, which states that within the kinetic control region, the gasification reaction rate should monotonically increase with increasing temperature. The calculation method is as follows: the partial derivative of the syngas component prediction vector with respect to the input gasification temperature is calculated using the automatic differentiation function of the deep learning framework. When this partial derivative is negative, it violates physical laws, and a positive penalty value is generated by correcting the linear unit function; this is the kinetic monotonicity constraint. The weighted sum of the above conservation law constraint and the kinetic monotonicity constraint yields the physical constraint loss function. The physical constraint loss function is then used for final constraint optimization training. This constraint optimization training is an additional fine-tuning training based on the non-equilibrium kinetic flow subnet trained in step S22, using the physical constraint loss function. In this training, the loss function is not the mean squared error, but rather a weighted sum of the mean squared error between the output of the residual two-stream network architecture and the actual output, plus the physical constraint loss function; this is the composite loss function. Using the same training and validation sets as in step S22, the partial derivatives of the composite loss function with respect to the network weight parameters within the non-equilibrium dynamic flow subnet are calculated using the backpropagation algorithm. The optimizer updates the weights until the composite loss function converges on the validation set to below a preset fine-tuning convergence threshold. This fine-tuning convergence threshold differs from the adversarial threshold; it serves as a standard to determine whether the current "fine-tuning" training process has reached a stable state and can be terminated early. It is set based on the initial composite loss function value calculated on the validation set before the fine-tuning training begins, setting a small value that indicates the composite loss function has entered a stable convergence phase. For example, it is set to 0.05% of the initial composite loss function value. Finally, a fine-tuned residual two-stream network architecture is obtained, defined as a physically constrained optimization model. See [link to relevant documentation]. Figure 3 This is a schematic diagram of the network weight parameter optimization trajectory under a multidimensional physical constraint loss manifold provided in an embodiment of the present invention. Figure 3As shown in the figure, this diagram schematically illustrates a two-dimensional cross-section of the network weight parameter space within the non-equilibrium dynamic flow subnet. The network weight parameter space is a high-dimensional space composed of all trainable network weight parameters in the non-equilibrium dynamic flow subnet. For ease of visualization, the horizontal axis parameter W1 and the vertical axis parameter W2 in the figure represent the two weight parameter dimensions in this high-dimensional space, respectively. The "initial point" in the figure represents the initial state of the network weight parameters within the non-equilibrium dynamic flow subnet before performing the fine-tuning training in step S31, i.e., the state after training in step S22. The blue surface in the figure represents the "physically constrained loss manifold" composed of all solutions that enable the output of the non-equilibrium dynamic flow subnet to satisfy the conservation law constraint and the dynamic monotonicity constraint. An optimization trajectory driven solely by mean square error, as shown by the dashed line, will deviate from this manifold in an attempt to simply reduce data error, eventually converging to a physically unreasonable solution, as shown by the unconstrained optimal solution in the figure. This invention incorporates a physical constraint loss function into the loss function to form a composite loss function. This generates a "pull" towards the physical constraint loss manifold during the optimization process, as shown by the red arrow in the figure. This forces the optimization trajectory driven by the composite loss function, as shown by the solid yellow line in the figure, to always search on the surface of the physical constraint loss manifold or its vicinity, eventually converging to a globally optimal solution that is both well-fitted by the data and physically consistent. The optimal solution of this invention is shown by the solid yellow dot in the figure.

[0094] Step S32: Perform reverse optimization on the physical constraint optimization model and output optimized operating parameters to guide the gasification process.

[0095] After obtaining the physical constraint optimization model, in order to transform it from a passive prediction tool into an optimization engine that can actively guide actual production, a gradient-based backpropagation optimization operation is executed. The goal is no longer to train or update the network weight parameters of the model, but to treat the physical constraint optimization model itself as a fixed, differentiable function and use the backpropagation algorithm to find the optimal combination of input operation parameters that enables the preset process objectives to be achieved.

[0096] Based on actual production needs, a process optimization objective function is defined. This objective function is a function that combines multiple process objectives into a single scalar value; the magnitude of the objective function reflects the quality of the current process conditions. For example, the objective function is constructed using the sum of the final predicted hydrogen gas fraction and the final predicted carbon monoxide volume fraction in the syngas component prediction vector as the yield index; and the predicted tar quality yield output by the tar and semi-coke prediction sub-models as the tar index. The yield index and the tar index are combined into the process optimization objective function through a linear weighted summation. After obtaining the process optimization objective function, a back-optimization calculation is performed. Specifically, all network weight parameters of the thermodynamic equilibrium flow sub-network and the non-equilibrium kinetic flow sub-network in the fixed physical constraint optimization model remain unchanged during the back-optimization calculation. An initial set of operating condition feature vectors is defined as the starting search point for the reverse optimization calculation. These vectors are set according to the actual production scenario, such as the operating parameters currently in use, like a gasification temperature of 900℃ and an air equivalence ratio of 0.3. The initial operating condition feature vectors are then concatenated with a set of fixed raw material feature vectors to form an initial feature tensor. These fixed raw material feature vectors are vectors that remain unchanged throughout the reverse optimization calculation, representing the physicochemical properties of the biomass raw material to be processed. The setting is based on the actual detection data of a batch or type of biomass raw material about to enter the gasifier. The initial operating condition feature vector is concatenated with the fixed raw material feature vector to form an initial feature tensor. This initial feature tensor is used as a fixed physical constraint optimization model to output an initial set of predicted final syngas components, defined as the initial syngas component prediction result set. This initial syngas component prediction result set is then input into the process optimization objective function to obtain its function value. The gradient of the function value with respect to each operable variable in the operating condition feature vector is calculated using a backpropagation algorithm. These operable variables are process parameters that operators can directly set and adjust during actual production, such as gasification temperature, gasification pressure, air equivalence ratio, and the water vapor to biomass mass ratio. The gradient indicates the direction in which each operable parameter should be adjusted to improve the process optimization objective function value. The initial operating condition feature vector is iteratively updated using optimization algorithms such as gradient ascent or gradient descent until the function value reaches its optimum or converges. Finally, the operating condition feature vector corresponding to the gradient is obtained, which serves as the optimization parameter to guide the gasification process. The optimized operating parameters are a set of specific target values ​​that can be directly used to set actual production equipment. For example, the recommended gasification temperature is 950℃ and the air equivalence ratio is 0.28. This provides a core technical means to achieve intelligent and optimized operation of the biomass gasification process.

[0097] Step S30, by constructing a physical constraint loss function and performing reverse optimization, solves the technical problem that the pre-constructed residual dual-stream network architecture may still output prediction results that violate physical laws when facing unknown operating conditions, and that the model itself can only passively predict and cannot actively guide production. This achieves a functional closed loop from high-fidelity prediction to proactive process optimization. Specifically, the fine-tuning training of the physical constraint loss function, by using conservation laws and dynamic monotonicity as mathematical penalty terms, endows the model with inherent physical consistency, resulting in the final physical constraint optimization model. Reverse optimization treats the physical constraint optimization model as a differentiable function and, through gradient calculation, derives the optimization operation parameters that enable the process optimization objective function to reach its optimum, transforming the prediction model into an intelligent process optimization engine.

[0098] Example 2

[0099] This embodiment, based on Embodiment 1, provides a biomass gasification reaction modeling system based on a physical information neural network, such as... Figure 4 As shown, it includes:

[0100] Data feature enhancement module: used to acquire raw gasification data characterizing the biomass gasification reaction process, perform data tensor quantization on the raw gasification data, output feature tensors for data enhancement processing, perform adversarial game training on the feature tensors to obtain an adversarial network model, sample the working condition area not covered by the experiment and call the adversarial network model to obtain an enhanced gasification feature dataset.

[0101] Hybrid model pre-construction module: used to perform thermodynamic equilibrium flow subnet calculation based on enhanced gasification feature dataset, obtain theoretical equilibrium prediction result set, construct and train non-equilibrium dynamic flow subnet through theoretical equilibrium prediction result set, call non-equilibrium dynamic flow subnet for calculation, obtain prediction dynamic deviation vector, perform fusion on prediction dynamic deviation vector, obtain final syngas component prediction result set to complete the pre-construction of residual two-flow network architecture;

[0102] Process optimization module: Used to construct a physical constraint loss function to fine-tune the residual two-stream network architecture, obtain a physical constraint optimization model, perform inverse optimization on the physical constraint optimization model, and output optimized operating parameters to guide the gasification process.

[0103] In the data feature enhancement module, the process involves acquiring raw gasification data characterizing the biomass gasification reaction process, performing data tensor quantization on the raw gasification data, outputting feature tensors for data enhancement processing, performing adversarial game training on the feature tensors to obtain an adversarial network model, sampling the operating conditions not covered by the experiment and calling the adversarial network model to obtain an enhanced gasification feature dataset, including:

[0104] Step S11: Obtain raw gasification data characterizing the biomass gasification reaction process, perform data tensor quantization on the raw gasification data, and output feature tensors for data augmentation processing.

[0105] Step S12: Construct a conditional generator and a convolutional discriminator based on the feature tensor, and perform adversarial game training on the conditional generator and the convolutional discriminator to obtain the adversarial network model.

[0106] Step S13: Sample the working area not covered by the experiment and call the adversarial network model to obtain the enhanced gasification feature dataset.

[0107] In the hybrid model pre-construction module, the thermodynamic equilibrium flow subnet calculation based on the enhanced gasification feature dataset is performed to obtain a theoretical equilibrium prediction result set. A non-equilibrium kinetic flow subnet is constructed and trained using this theoretical equilibrium prediction result set. The non-equilibrium kinetic flow subnet is then called for calculation to obtain the predicted kinetic deviation vector. Finally, the predicted kinetic deviation vector is fused to obtain the final syngas component prediction result set, thus completing the pre-construction of the residual two-flow network architecture. This includes:

[0108] Step S21: Perform thermodynamic equilibrium flow subnet calculation based on the enhanced gasification feature dataset to obtain the theoretical equilibrium state prediction result set;

[0109] Step S22: Construct and train a non-equilibrium dynamic subnet based on the theoretical equilibrium prediction result set;

[0110] Step S23: The non-equilibrium dynamic flow subnet is called for calculation to obtain the predicted dynamic deviation vector. The predicted dynamic deviation vector is fused to obtain the final syngas component prediction result set, thus completing the pre-construction of the residual two-flow network architecture.

[0111] In the process optimization module, the physical constraint loss function is constructed to fine-tune the residual two-stream network architecture to obtain a physical constraint optimization model. Backward optimization is then performed on the physical constraint optimization model to output optimized operating parameters for guiding the gasification process, including:

[0112] Step S31: Construct a physical constraint loss function, and use the physical constraint loss function to fine-tune and train the residual two-stream network architecture to obtain a physical constraint optimization model;

[0113] Step S32: Perform reverse optimization on the physical constraint optimization model and output optimized operating parameters to guide the gasification process.

[0114] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0115] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A biomass gasification reaction modeling method based on physical information neural networks, characterized in that, The method includes: The raw gasification data characterizing the biomass gasification reaction process is obtained. Data tensor operation is performed on the raw gasification data to output the feature tensor for data augmentation. Adversarial game training is performed on the feature tensor to obtain the adversarial network model. The working condition area not covered by the experiment is sampled and the adversarial network model is called to obtain the enhanced gasification feature dataset. Thermodynamic equilibrium flow subnet calculations are performed based on the enhanced gasification feature dataset to obtain a theoretical equilibrium prediction result set. A non-equilibrium dynamic flow subnet is constructed and trained using the theoretical equilibrium prediction result set. The non-equilibrium dynamic flow subnet is called for calculation to obtain the predicted dynamic deviation vector. The predicted dynamic deviation vector is fused to obtain the final syngas component prediction result set, thus completing the pre-construction of the residual two-flow network architecture. A physical constraint loss function is constructed to fine-tune the residual two-stream network architecture, resulting in a physical constraint optimization model. Inverse optimization is then performed on the physical constraint optimization model to output optimized operating parameters for guiding the gasification process. The method for obtaining the theoretical equilibrium state prediction result set includes: the enhanced gasification feature dataset includes the simulation source feature tensor, the experimental source feature tensor, and the enhanced feature tensor of the region not covered by the experiment; For each feature tensor in the enhanced gasification feature dataset, atomic composition information, gasification temperature, and gasification pressure are extracted from the working condition feature vector and raw material feature vector contained therein as input parameters. The input parameters are fed into a thermodynamic equilibrium subnet for calculation, resulting in a set of solutions describing the molar number of each gas phase product under equilibrium conditions. The solutions are converted into their respective volume fractions in the total gas to obtain the theoretical equilibrium prediction vector. The total gas refers to the mixture of all products in the reaction system that exist in gaseous form when thermodynamic equilibrium is reached. The theoretical equilibrium prediction vectors corresponding to each feature tensor in the enhanced gasification feature dataset are combined to obtain the theoretical equilibrium prediction result set. The method for obtaining the final syngas component prediction result set includes: for each feature tensor in the enhanced gasification feature dataset, calling the non-equilibrium dynamic flow subnet to calculate the prediction dynamic deviation vector corresponding to the feature tensor; The predicted dynamics deviation vector and the theoretical equilibrium state prediction vector corresponding to the feature tensor are added element by element to obtain a syngas component prediction vector. Each element value in the syngas component prediction vector is the result of adding the corresponding element value in the theoretical equilibrium state prediction vector and the corresponding element value in the predicted dynamics deviation vector. The syngas component prediction vectors calculated from all feature tensors are combined to obtain the final syngas component prediction result set.

2. The biomass gasification reaction modeling method based on physical information neural networks as described in claim 1, characterized in that, The method for obtaining the adversarial network model includes: The feature tensor includes experimental source feature tensors derived from experimental log files and simulation source feature tensors derived from simulation log files. The feature tensor is obtained by concatenating the raw material feature vector and the operating condition feature vector, which are obtained by performing normalization processing on the original gasification data. Construct an adversarial network model that includes a condition generator and a convolutional discriminator, wherein the network structure of the condition generator is a deep residual network and the network structure of the convolutional discriminator is a multi-layer convolutional neural network. After constructing the condition generator and the convolutional discriminator, adversarial game training is performed. This training alternately optimizes the convolutional discriminator and the condition generator until the total loss function converges to below the preset adversarial threshold, thereby outputting the adversarial network model.

3. The biomass gasification reaction modeling method based on physical information neural networks as described in claim 2, characterized in that, The optimization of the convolutional discriminant includes: Treating the network weight parameters inside the condition generator as constants, a convolutional discriminator loss function is calculated. The convolutional discriminator loss function is obtained by performing algebraic operations on the scalar values ​​corresponding to the virtual high-fidelity feature tensor generated by the condition generator and the scalar values ​​corresponding to the experimental source feature tensor. The partial derivatives of the convolutional discriminator loss function with respect to the internal network weight parameters of the convolutional discriminator are calculated using the backpropagation algorithm, and the weights are updated using an optimizer to minimize the convolutional discriminator loss function.

4. The biomass gasification reaction modeling method based on physical information neural networks as described in claim 3, characterized in that, The non-equilibrium dynamic flow subnet includes: For each feature tensor in the enhanced gasification feature dataset, extract the corresponding value of the feature tensor from the corresponding original gasification data record to form a target state vector. Subtract the target state vector from the theoretical equilibrium prediction vector corresponding to the feature tensor element by element to obtain a target deviation vector. Combine the target deviation vectors corresponding to each feature tensor to obtain the target deviation dataset. The enhanced gasification feature dataset and the corresponding target deviation dataset are divided into a training set and a validation set. In each training iteration cycle, a training batch is extracted from the training set, and the joint input vector in the training batch is input into the non-equilibrium dynamic flow subnet to obtain the prediction deviation batch. The mean square error between the prediction deviation batch and the target deviation vector is calculated as the loss function, and the network weights are updated through backpropagation until the loss function converges on the validation set to obtain the non-equilibrium dynamic flow subnet.

5. The biomass gasification reaction modeling method based on physical information neural networks as described in claim 4, characterized in that, The physical constraint loss function includes: The physical constraint loss function consists of two parts. The first part is the conservation law constraint term, which includes the element mass conservation residual and the system enthalpy change residual. The element mass conservation residual is obtained by calculating the absolute value of the difference between the initial total element mass of the reaction system and the predicted total element mass of the product. The initial total element mass of the reaction system is calculated based on the raw material feature vector and the operating condition feature vector in the feature tensor. The predicted total element mass of the product is calculated based on the residual two-stream network architecture. The system enthalpy change residual is obtained by calling a thermodynamic database. The second part is the kinetic monotonicity constraint term, which is obtained by calculating the partial derivative of the syngas component prediction vector with respect to the input gasification temperature, and generating a positive penalty value when the partial derivative is negative. The physical constraint loss function is obtained by weighted summation of the conservation law constraint term and the dynamic monotonicity constraint term.

6. The biomass gasification reaction modeling method based on physical information neural networks as described in claim 5, characterized in that, The method for obtaining the physical constraint optimization model includes: A composite loss function is constructed by weighted summation of two components. The first part is the data fitting error term obtained by calculating the mean square error between the syngas component prediction vector output by the residual two-stream network architecture and the actual output result corresponding to the feature tensor. The second part is the physical constraint loss function. Using the composite loss function, the partial derivatives of the composite loss function with respect to the network weight parameters inside the non-equilibrium dynamic flow subnet are calculated through the backpropagation algorithm. The weights are then updated using an optimizer until the value of the composite loss function converges to below a preset fine-tuning convergence threshold, thus obtaining the physical constraint optimization model.

7. The biomass gasification reaction modeling method based on physical information neural networks as described in claim 6, characterized in that, The method for obtaining the optimized operation parameters includes: Define a process optimization objective function, which takes the output of the physical constraint optimization model as input, and outputs a scalar value for evaluating the merits of the current process conditions by performing a preset mathematical combination operation on the input. Fix all network weight parameters of the physical constraint optimization model, and set a set of initial working condition feature vectors and a set of fixed raw material feature vectors; The gradient of the process optimization objective function with respect to each operable variable in the operating condition feature vector is calculated using the backpropagation algorithm. This gradient is then used to iteratively update the initial operating condition feature vector until the value of the process optimization objective function no longer improves after several consecutive iterations. At this point, the optimal value is determined, and the operating condition feature vector at this point is used as the optimization operation parameter. The operable variables are the process parameters that constitute the operating condition feature vector and can be directly set in the actual production process.

8. A biomass gasification reaction modeling system based on a physical information neural network, used to implement the biomass gasification reaction modeling method based on a physical information neural network as described in any one of claims 1-7, characterized in that, The system includes: Data feature enhancement module: used to acquire raw gasification data characterizing the biomass gasification reaction process, perform data tensor quantization on the raw gasification data, output feature tensors for data enhancement processing, perform adversarial game training on the feature tensors to obtain an adversarial network model, sample the working condition area not covered by the experiment and call the adversarial network model to obtain an enhanced gasification feature dataset. Hybrid model pre-construction module: used to perform thermodynamic equilibrium flow subnet calculation based on enhanced gasification feature dataset, obtain theoretical equilibrium prediction result set, construct and train non-equilibrium dynamic flow subnet through theoretical equilibrium prediction result set, call non-equilibrium dynamic flow subnet for calculation, obtain prediction dynamic deviation vector, perform fusion on prediction dynamic deviation vector, obtain final syngas component prediction result set to complete the pre-construction of residual two-flow network architecture; Process optimization module: Used to construct a physical constraint loss function to fine-tune the residual two-stream network architecture, obtain a physical constraint optimization model, perform inverse optimization on the physical constraint optimization model, and output optimized operating parameters to guide the gasification process.

Images