CN-121787287-B - Biomass gasification reaction modeling method and system based on physical information neural network

CN121787287BCN 121787287 BCN121787287 BCN 121787287BCN-121787287-B

Abstract

The invention 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 steps of obtaining original gasification data and executing data Zhang Lianghua to obtain a characteristic tensor, training an countermeasure network model to generate an enhanced gasification characteristic data set covering unobserved working conditions, constructing a residual double-flow network architecture including thermodynamic equilibrium flow sub-network calculation theoretical basis and unbalanced dynamic flow sub-network learning deviation correction, constructing a physical constraint loss function to conduct fine tuning training on the residual double-flow network architecture to obtain a physical constraint optimization model, and executing reverse optimization on the physical constraint optimization model to output optimization operation parameters. The method solves the problems of prediction deviation and violation of physical rules of the traditional model, realizes a functional closed loop from high-precision prediction to active process optimization, and remarkably improves the prediction precision and generalization capability of the model under raw material fluctuation.

Inventors

LI CUNLEI
FAN QINSHAN
FU YIWEI
Cheng Manqiu
ZHENG SHIJIN
DING HONG
ZHOU FEI
ZHANG PENG
Ye Xingpei
LIU MINGRUI
LIU ZHOU
Jie Xiaochen

Assignees

江苏省国信研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20260305

Claims (8)

1. The biomass gasification reaction modeling method based on the physical information neural network is characterized by comprising the following steps of: Acquiring original gasification data representing a biomass gasification reaction process, performing data tensioning operation on the original gasification data, outputting a characteristic tensor for data enhancement processing, performing countermeasure game training on the characteristic tensor to obtain a countermeasure network model, sampling a working condition area which is not covered by an experiment, and calling the countermeasure network model to obtain an enhanced gasification characteristic data set; Performing thermodynamic equilibrium flow sub-network calculation based on the enhanced gasification characteristic data set to obtain a theoretical equilibrium state prediction result set, constructing and training an unbalanced dynamic flow sub-network through the theoretical equilibrium state prediction result set, calling the unbalanced dynamic flow sub-network to perform calculation to obtain a prediction dynamic deviation vector, and performing fusion on the prediction dynamic deviation vector to obtain a final synthesis gas component prediction result set to finish the pre-construction of a residual double-flow network architecture; Constructing a physical constraint loss function to perform fine tuning training on the residual double-flow network architecture to obtain a physical constraint optimization model, performing reverse optimization on the physical constraint optimization model, and outputting optimized operation parameters for guiding the gasification process; The method for acquiring the theoretical equilibrium state prediction result set comprises the steps that an enhanced gasification characteristic data set comprises a simulation source characteristic tensor and an experimental source characteristic tensor, and the enhanced characteristic tensor of an area which is not covered by an experiment; For each characteristic tensor in the enhanced gasification characteristic data set, extracting atomic composition information, gasification temperature and gasification pressure from the working condition characteristic vector and the raw material characteristic vector contained in the characteristic tensor as input parameters; sending the input parameters into a thermodynamic equilibrium flow sub-network for calculation to obtain a group of solutions describing the mole numbers of each gas phase product in an equilibrium state; converting the solution into volume fractions of the total gas to obtain theoretical equilibrium state prediction vectors, wherein the total gas refers to a mixture of all products existing in a gaseous form in a reaction system when thermodynamic equilibrium is reached, and combining the theoretical equilibrium state prediction vectors corresponding to each characteristic tensor in the enhanced gasification characteristic data set to obtain a theoretical equilibrium state prediction result set; The method for obtaining the final synthesis gas component prediction result set comprises the steps of calling an unbalanced dynamic flow subnet for each characteristic tensor in the enhanced gasification characteristic data set, and calculating a prediction dynamic deviation vector corresponding to the characteristic tensor; performing element-by-element addition operation on the predicted dynamic deviation vector and a theoretical equilibrium state predicted vector corresponding to the characteristic tensor to obtain a synthesis gas component predicted vector, wherein each element value in the synthesis gas component predicted vector is the result of adding a corresponding element value in the theoretical equilibrium state predicted vector and a corresponding element value in the predicted dynamic deviation vector; and combining the synthesis gas component prediction vectors calculated by all the characteristic tensors to obtain a final synthesis gas component prediction result set.
2. The physical information neural network-based biomass gasification reaction modeling method as claimed in claim 1, wherein the method for obtaining the countermeasure network model comprises the steps of: the characteristic tensor comprises an experimental source characteristic tensor from an experimental log file and a simulation source characteristic tensor from a simulation log file, wherein the characteristic tensor is obtained by splicing a raw material characteristic vector and a working condition characteristic vector, and the raw material characteristic vector and the working condition characteristic vector are obtained by carrying out normalization processing on original gasification data; Constructing an countermeasure network model comprising a condition generator and a convolution discriminator, wherein the network structure of the condition generator adopts a depth residual error network, and the network structure of the convolution discriminator adopts a multi-layer convolution neural network; After the condition generator and the convolution arbiter are constructed, an countermeasure game training is performed, wherein the training is performed by alternately performing the convolution arbiter optimization and the condition generator optimization until the total loss function converges below a preset countermeasure threshold, thereby outputting the countermeasure network model.
3. The physical information neural network-based biomass gasification reaction modeling method as claimed in claim 2, wherein the convolution discriminator optimization includes: Taking the network weight parameter in the condition generator as a constant, and calculating a convolution discriminant loss function, wherein the convolution discriminant loss function is obtained by generating a virtual high-fidelity characteristic tensor according to the condition generator and performing algebraic operation on a scalar value corresponding to the virtual high-fidelity characteristic tensor and a scalar value corresponding to the experimental source characteristic tensor; And calculating the partial derivative of the convolution discriminant loss function to the internal network weight parameter of the convolution discriminant through a back propagation algorithm, and updating the weight of the partial derivative by using an optimizer so as to minimize the convolution discriminant loss function.
4. A physical information neural network-based biomass gasification reaction modeling method according to claim 3, wherein said unbalanced dynamic flow subnetwork comprises: for each characteristic tensor in the enhanced gasification characteristic data set, extracting a numerical value corresponding to the characteristic tensor from the corresponding original gasification data record to form a target state vector, subtracting the target state vector from the theoretical equilibrium state prediction vector corresponding to the characteristic tensor element by element to obtain a target deviation vector, and combining the target deviation vectors corresponding to each characteristic tensor to obtain a target deviation data set; Dividing an enhanced gasification characteristic data set and a corresponding target deviation data set into a training set and a verification set, extracting a training batch from the training set in each training iteration period, inputting a combined input vector in the training batch to an unbalanced dynamic flow sub-network to obtain a predicted deviation batch, calculating a mean square error between the predicted deviation batch and the target deviation vector as a loss function, and updating a network weight through counter propagation until the loss function converges on the verification set to obtain the unbalanced dynamic flow sub-network.
5. The physical information neural network-based biomass gasification reaction modeling method of claim 4, wherein the physical constraint loss function comprises: The physical constraint loss function comprises two parts, wherein the first part is a conservation law constraint term which comprises an element mass conservation residual error and a system enthalpy change residual error, the element mass conservation residual error is obtained by calculating the absolute value of the difference between the initial total element mass of a reaction system and the total element mass of a predicted product, the initial total element mass of the reaction system is calculated according to a raw material characteristic vector and a working condition characteristic vector in a characteristic tensor, the total element mass of the predicted product is calculated according to a residual double-current network architecture, and the system enthalpy change residual error is obtained by calling a thermodynamic database; The second part is a dynamics monotonicity constraint term which is obtained by calculating a partial derivative of the syngas component predictive vector with respect to the input gasification temperature and generating a positive penalty value when the partial derivative is negative; And carrying out weighted summation on the conservation law constraint term and the dynamics monotonicity constraint term to obtain a physical constraint loss function.
6. The biomass gasification reaction modeling method based on the physical information neural network as claimed in claim 5, wherein the method for obtaining the physical constraint optimization model comprises the following steps: Constructing a composite loss function, wherein the composite loss function is formed by carrying out weighted summation on two components, the first component is a data fitting error term obtained by calculating the mean square error between a predicted vector of a synthesis gas component output by a residual double-flow network architecture and a real output result corresponding to a characteristic tensor, and the second component is the physical constraint loss function; And calculating the partial derivative of the composite loss function to the internal network weight parameter of the unbalanced dynamic flow sub-network by using the composite loss function through a back propagation algorithm, and updating the weight by using an optimizer until the numerical value of the composite loss function is converged below a preset fine tuning convergence threshold value to obtain the physical constraint optimization model.
7. The physical information neural network-based biomass gasification reaction modeling method as claimed in claim 6, wherein the method for obtaining the optimized operation parameters comprises: Defining a process optimization objective function, wherein the process optimization objective function takes the output of the physical constraint optimization model as input, and outputs a scalar value for evaluating the quality of the current process condition by executing a preset mathematical combination operation on the input; Fixing all network weight parameters of the physical constraint optimization model, and setting a group of initial working condition feature vectors and a group of fixed raw material feature vectors; calculating the gradient of each operable variable in the working condition feature vector by the process optimization objective function through a back propagation algorithm, and carrying out iterative updating on the initial working condition feature vector by utilizing the gradient until the value of the process optimization objective function is judged to be optimal when a plurality of continuous iterative periods are not improved any more, and taking the working condition feature vector at the moment as an optimized operation parameter, wherein the operable variable is a process parameter which forms the working condition feature vector and can be directly set in the actual production process.
8. A physical information neural network-based biomass gasification reaction modeling system for implementing the physical information neural network-based biomass gasification reaction modeling method according to any one of claims 1 to 7, characterized in that the system comprises: The data characteristic enhancement module is used for acquiring original gasification data representing a biomass gasification reaction process, executing data tensioning operation on the original gasification data, outputting characteristic tensors for data enhancement processing, executing countermeasure game training on the characteristic tensors to obtain a countermeasure network model, sampling a working condition area which is not covered by an experiment, and calling the countermeasure network model to obtain an enhanced gasification characteristic data set; The mixed model pre-construction module is used for executing thermodynamic equilibrium flow sub-network calculation based on the enhanced gasification characteristic data set to obtain a theoretical equilibrium state prediction result set, constructing and training an unbalanced dynamic flow sub-network through the theoretical equilibrium state prediction result set, calling the unbalanced dynamic flow sub-network to calculate to obtain a prediction dynamics deviation vector, and executing fusion on the prediction dynamics deviation vector to obtain a final synthesis gas component prediction result set to finish pre-construction of a residual double-flow network architecture; The process optimizing module is used for constructing a physical constraint loss function to perform fine tuning training on the residual double-flow network architecture to obtain a physical constraint optimizing model, performing reverse optimizing on the physical constraint optimizing model and outputting optimized operation parameters for guiding the gasification process.

Description

Biomass gasification reaction modeling method and system based on physical information neural network Technical Field The invention relates to the technical field of biomass energy process modeling, in particular to a biomass gasification reaction modeling method and system based on a physical information neural network. Background Along with the increasing importance of biomass gasification technology in energy transformation, the prediction accuracy bottleneck of a process control model becomes a key for restricting the development of industry. When the traditional model is used for coping with raw material fluctuation or working condition change, the deviation between a predicted result and an actual value is huge, and the stability of a downstream process is seriously affected. The prior art has two polarization limitations that obvious unbalanced kinetic effects such as tar generation, incomplete carbon conversion and the like in the gasification process are ignored based on a thermodynamic equilibrium mechanism model, so that prediction is severely distorted, and a pure data-driven machine learning model is easy to overfill when facing sparse and high-dimensional data of an industrial field, and often outputs paradoxical results violating basic physical laws such as quality, energy conservation and the like. The traditional mixed modeling method stays in simple linear correction, and cannot realize deep fusion of mechanism and data. Therefore, how to combine the explicit physical law as hard constraint and the nonlinear fitting capability of data driving to construct a modeling framework which can not only follow the physical iron law, but also adaptively correct the unbalanced state deviation so as to realize high-precision and interpretable prediction under the high-dimensional and sparse data condition is a technical problem to be solved urgently in the field. In the prior art, china patent with the authority of publication number CN117008479B discloses a carbon emission optimizing control method and system based on a biomass gasification furnace. The system comprises a data preprocessing unit, a parameter feature classification module, a distributed feature extraction unit, a strategy network calculation module and a gasifier control algorithm module, and can realize functions of gasification data classification, distributed feature extraction, dynamic model construction, operation parameter optimization and the like. According to the method, a dynamic analysis model of the gasification reaction is built through an LSTM network and a full-connection network, the accuracy of biomass gasification parameter prediction is improved, and data support is provided for gasifier operation parameter adjustment. Chinese patent application publication No. CN120409193a discloses a biomass gasification process twin optimization system and method thereof. The system adopts a multi-module collaborative architecture, comprises multi-source data intelligent perception, hybrid evolution algorithm self-optimization, knowledge-graph self-adaptive reconstruction, multi-time-scale collaborative calibration and prediction and closed-loop feedback iteration symbiotic modules, and respectively realizes the functions of model deviation recognition, parameter optimization, model reconstruction, accurate prediction and iteration optimization. The method combines a digital twin technology with an evolutionary algorithm to realize the transformation of the gasification process from passive early warning to active optimization, and improves the long-term consistency and adaptability of the model. However, although the two prior arts have a certain value in aspects of gasification data processing, model optimization and process control, the core pain point of cooperation of physical law constraint and unbalanced state deviation adaptive correction in the current biomass gasification modeling cannot be solved. The patent with the issued publication number of CN117008479B focuses on dynamic prediction driven by data, does not integrate physical laws such as mass conservation, energy conservation and the like into a model framework as hard constraint, is easy to output a result violating basic physical rules, and lacks anti-overfitting design when facing sparse high-dimensional data. The combination of the focusing digital twinning and the algorithm optimization of the patent with the publication number of CN120409193A does not realize the deep fusion of the mechanism knowledge and the data driving capability, and has limited correction effects on unbalanced kinetic effects such as tar generation, incomplete carbon conversion and the like in the gasification process. Neither does the modeling architecture with physical constraint and nonlinear fitting cooperation be constructed, the prediction precision can not be maintained when the raw materials fluctuate or the working conditions change, the interpretability and the self-adaptive