Zero-carbon power system modeling method and system based on closed-loop feedback optimization order reduction

Abstract

The invention discloses a zero-carbon power system modeling method and system based on closed-loop feedback optimization order reduction, and relates to the field of zero-carbon power system modeling and combustion mechanism optimizing.The method comprises the steps that mechanism order reduction is conducted on a chemical reaction mechanism network of a zero-carbon power system, and a key reaction path is obtained; performing dynamic layout on the energy network constructed based on the key reaction path by adopting an improved dynamic force model to obtain a dynamic energy network; optimizing the dynamic energy network by adopting a closed-loop feedback mechanism to obtain a simplified mechanism network; obtaining multi-dimensional data based on the simplified mechanism network, and performing feature extraction to obtain a low-dimensional feature vector; and inputting the low-dimensional feature vectors into the comprehensive mapping model for prediction to obtain a prediction result of the key combustion parameters. A directed relation graph method of plasma with error propagation is combined with a closed-loop feedback mechanism to reduce the order of a mechanism network, dimension reduction is performed on multi-dimensional data of the mechanism network, and high-precision and quick response prediction of key combustion parameters is realized by utilizing a multi-model fusion strategy.

Description

Technical Field

[0001] The present invention relates to the field of zero-carbon power system modeling and combustion mechanism optimization, and particularly to a zero-carbon power system modeling method and system based on closed-loop feedback optimization and order reduction. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] With the increasing global demand for clean energy and efficient combustion technologies, zero-carbon power systems have become an important direction for future energy transformation. With the global high emphasis on zero carbon emissions and sustainable energy, zero-carbon power systems have become a key direction for energy technology innovation. Traditional combustion technologies are gradually difficult to meet the requirements of the new generation of clean energy systems in terms of energy efficiency, emission control, and system safety. Therefore, high-precision modeling and real-time optimization methods based on digital technologies have received extensive attention. Digital zero-carbon power systems can not only achieve efficient energy conversion but also ensure the stable and safe operation of the system under complex working conditions through the integration of high-precision modeling, real-time monitoring, and intelligent optimization.

[0004] In recent years, plasma-assisted combustion technology has received attention because it can stimulate combustion reactions at low temperatures, accelerate fuel decomposition, and expand the fuel flammability limit. However, it also brings complex non-linear and multi-scale coupling problems. Traditional combustion mechanisms mostly rely on thermochemical reaction descriptions and are difficult to effectively solve the high computational complexity and time-scale stiffness problems caused by microscopic processes such as electron collision, excitation, and ionization during plasma discharge. These challenges not only limit the possibility of real-time simulation and optimization of the combustion process but also pose higher requirements for the digital modeling and control of zero-carbon power systems.

[0005] Currently, researchers have tried to simplify the mechanism of complex reaction networks using global path analysis, zero-dimensional solvers, and data-driven dimensionality reduction methods. However, most methods still have problems such as limited applicability, difficulty in balancing computational accuracy and response speed, and are difficult to meet the requirements of future zero-carbon power systems for high-precision and fast response of the combustion process under multi-fuel and multi-condition conditions. For example, many simplification methods ignore the multi-scale coupling effects occurring in complex combustion processes, especially the interaction between plasma and fuel systems. Due to the strong non-linear characteristics of processes such as electron collision, excitation, and ionization in plasma, traditional methods usually cannot provide sufficient real-time response and accuracy when dealing with these microscopic processes. Summary of the Invention

[0006] To overcome the deficiencies of the above-mentioned existing technologies and address the problems of complex plasma-assisted combustion reaction mechanism networks, low multi-scale coupling calculation efficiency, and insufficient real-time control accuracy in zero-carbon power systems, the present invention provides a zero-carbon power system modeling method and system based on closed-loop feedback optimization and order reduction. By reducing the dimensionality of the reaction mechanism network and input data while retaining key physical and chemical processes, and using a multi-model fusion strategy to capture the non-linear mapping between key variables and predicted outputs, the efficiency and accuracy of numerical simulations are significantly improved.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides a zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction, including: Obtain the chemical reaction mechanism network and its multi-dimensional data of the zero-carbon power system; Perform mechanism order reduction on the chemical reaction mechanism network based on the multi-dimensional data to obtain key reaction paths; Construct an energy network based on the key reaction paths, perform dynamic layout on the energy network using an improved dynamic force model to obtain a dynamic energy network; optimize the dynamic energy network using a closed-loop feedback mechanism to obtain a simplified mechanism network; Obtain the multi-dimensional data of the simplified mechanism network and perform preprocessing, perform principal component analysis on the preprocessed multi-dimensional data to obtain low-dimensional feature vectors; Input the low-dimensional feature vectors into a pre-trained comprehensive mapping model for prediction to obtain the prediction results of key combustion parameters.

[0008] In a further technical solution, the mechanism order reduction of the chemical reaction mechanism network is performed using a directed relationship graph method with plasma error propagation, calculate the direct interaction coefficients between species and various energy branch variables, and when the direct interaction coefficient is less than a set error threshold, eliminate the corresponding species and their related redundant reactions to obtain a skeletal mechanism.

[0009] In a further technical solution, first use depth-first search to traverse the skeletal mechanism to automatically identify key reaction paths, and then perform a review of the retained reaction paths in combination with reaction path analysis.

[0010] In a further technical solution, constructing an energy network based on key reaction paths specifically means: constructing a visualization network based on key reaction paths, and stratifying different species in the visualization network based on the energy types of the species to obtain an energy network.

[0011] In a further technical solution, an energy-sensitive force mechanism is introduced into the ForceAtlas2 model to obtain an improved dynamic force model, including a repulsive force model, an attractive force model, and a random perturbation term.

[0012] A further technical solution is that the closed-loop feedback mechanism is specifically as follows: Use an improved dynamic force model to perform spatial layout on the energy network, visually sort the species nodes according to the reaction intensity and influence, and obtain the node importance ranking; Calibrate the high-importance nodes based on the node importance ranking, and calculate the average distance between each pair according to the spatial coordinates of the high-importance nodes; Based on the average distance, use a feedback adjustment strategy to adjust the error threshold in the directed graph method with plasma band error propagation, and perform mechanism order reduction based on the adjusted error threshold; Iterate the above steps. When the error thresholds of all species nodes meet the convergence conditions, stop the closed-loop and output the simplified mechanism network.

[0013] A further technical solution is to form a comprehensive mapping model based on a multi-model fusion prediction framework, specifically as follows: Construct a candidate model library containing multiple candidate models; Use all candidate models to perform training and performance evaluation in the way of k-fold cross-validation, and obtain the fusion weights of each candidate model based on the performance evaluation; Construct a comprehensive mapping model based on each candidate model and the corresponding fusion weights.

[0014] In a second aspect, the present invention provides a zero-carbon power system modeling system based on closed-loop feedback optimization and order reduction, including: A data acquisition module, which is configured to: acquire the chemical reaction mechanism network and its multi-dimensional data of the zero-carbon power system; A mechanism order reduction module, which is configured to: perform mechanism order reduction on the chemical reaction mechanism network based on the multi-dimensional data to obtain the key reaction path; A closed-loop feedback optimization module, which is configured to: construct an energy network based on the key reaction path, perform dynamic layout on the energy network using an improved dynamic force model to obtain a dynamic energy network; use a closed-loop feedback mechanism to optimize the dynamic energy network to obtain a simplified mechanism network; A feature extraction module, which is configured to: acquire the multi-dimensional data of the simplified mechanism network and perform preprocessing, and perform principal component analysis on the preprocessed multi-dimensional data to obtain a low-dimensional feature vector; A model prediction module, which is configured to: input the low-dimensional feature vector into a pre-trained comprehensive mapping model for prediction to obtain the prediction result of the key combustion parameters.

[0015] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps in the zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as described in the first aspect are implemented.

[0016] In a fourth aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps in the zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as described in the first aspect are implemented.

[0017] The above one or more technical solutions have the following beneficial effects: Through the plasma band error propagation directed relationship graph method P-DRGEP, depth-first search method DFS, and reaction path analysis RPA, the present invention performs preliminary mechanism order reduction on the original chemical reaction mechanism network to obtain key reaction paths, and then uses an improved dynamic force model to realize the real-time visualization of the mechanism network. According to the visualization analysis of the mechanism network and combined with the closed-loop feedback mechanism, further optimization and order reduction of the key reaction paths are carried out to obtain the final simplified mechanism network, realizing the real-time optimization of the mechanism network and significantly improving the reliability of the mechanism simplification process.

[0018] The present invention reduces the dimensionality of multi-dimensional data through principal component analysis, extracts low-dimensional feature vectors, and combines them with a multi-model fusion prediction framework for prediction, breaking through the limitations of a single model and improving the prediction accuracy of key combustion parameters.

[0019] The present invention innovatively proposes a closed-loop feedback mechanism, which forms an intelligent closed loop through the dynamic adjustment of visualization layout and error threshold, ensuring the reliability and energy conservation of mechanism network simplification.

[0020] The present invention uses mechanism order reduction and data-driven multi-model integrated prediction to efficiently model and simulate and optimize complex chemical reaction processes. During the mechanism order reduction process, key physical and chemical processes are retained to ensure the accurate characterization of reaction kinetics and energy balance. At the same time, relying on advanced multi-model fusion prediction technology, traditional regression methods and modern machine learning algorithms are integrated to achieve high-precision and fast-response prediction of key combustion parameters (such as temperature, main species concentration, etc.) of the zero-carbon power system to meet the strict requirements for system safety, energy efficiency, and emission control under actual working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0022] Figure 1It is a flowchart of the zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction according to an embodiment of the present invention; Figure 2 It is a simplified flowchart of the visualization intelligent reaction mechanism according to an embodiment of the present invention. Detailed implementation manners

[0023] It should be noted that the following detailed descriptions are all exemplary and are intended to provide further explanations of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0024] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0026] Embodiment 1 As Figure 1 shown, this embodiment discloses a zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction, and the method includes the following steps: S1: Obtain the chemical reaction mechanism network and its multi-dimensional data of the zero-carbon power system; In this embodiment, for the research on the zero-carbon power system (such as plasma-assisted combustion in an / air-fuel system, etc.), first, a complete chemical reaction mechanism network needs to be constructed. This mechanism is based on the reaction data disclosed in existing literature and combines the plasma-fuel interaction process concerned by the present invention, and forms an input format applicable to solvers such as ZDPlasKin / ChemKin through screening, integration, and format conversion. The chemical reaction mechanism network includes key reaction paths such as electron collision, excitation, ionization, radical generation, fuel oxidation, and intermediate species consumption.

[0027] Among them, the reaction rate constants of each reaction are expressed by the Arrhenius formula:

[0028] Among them, represents the pre-exponential factor, represents the temperature, represents the temperature exponent, represents the activation energy, represents the gas constant.

[0029] The chemical reaction mechanism network details the generation, transformation, and consumption processes of each reaction path and intermediate products. To ensure the completeness of the mechanism data, it is necessary to collect multi-dimensional data such as temperature, pressure, species concentration, and reaction rate simultaneously, and preliminarily organize the collected multi-dimensional data (numerical simulation data) to provide basic support for subsequent data analysis.

[0030] S2: Mechanism reduction of the chemical reaction mechanism network based on the multi-dimensional data to obtain the key reaction paths; In this embodiment, since the complete mechanism contains a large number of reactions and species, directly solving often faces problems such as high computational complexity and difficulty in meeting real-time simulation. Therefore, the plasma directed relation graph method with error propagation (P-DRGEP), depth-first search method, and reaction path analysis method are used to reduce the order of the mechanism network.

[0031] Plasma directed relation graph method with error propagation (P-DRGEP): The DRG (Directed Reduced Graph) algorithm defines the species For the species The normalized contribution to the generation amount , thereby quantifying the species The dependence on the species If is very large, it means that removing the species from the detailed mechanism will cause a large error in the prediction of the species . Set an appropriate threshold. If is less than this threshold, it is considered that the species has a very small normalized contribution to the generation rate of the species , and the species and its related redundant reactions can be removed. If is greater than this threshold, the species is retained. By continuously adjusting the threshold, a skeletal mechanism with gradually fewer species and reactions is obtained.

[0032] Is defined as follows:

[0033] Among them, Represents the th reaction in the reaction sequence, Represents the elementary reaction series, Represents the th reaction, the net chemical reaction stoichiometric coefficient of the species , Represents the th reaction net chemical reaction rate; when and participate in the th reaction simultaneously. If so, the value is 1; otherwise, it is 0.

[0034] Error propagation is introduced into the DRG algorithm to obtain the directed relation graph method based on error propagation (DRGEP). The definition is as follows:

[0035]

[0036]

[0037] where represents the direct interaction relationship between species and species ; represents the reaction; represents the total generation rate of species , which is non - negative; represents the total digestion rate of species , which is non - negative.

[0038] The plasma - targeted directed relation graph method with error propagation P - DRGEP (Plasma - targeted Directed Relation Graph with Error Propagation) is obtained by introducing direct interactions of energy branches and strict error propagation analysis on the basis of traditional DRGEP. During the plasma - assisted combustion process, electrons lose energy in collision reactions such as ionization, electron excitation, vibrational excitation, and dissociation. These processes constitute the "energy branches", and their characteristics are different from those of a single chemical species in the traditional mechanism.

[0039] The direct interaction coefficient between species and the th energy - branch variable is denoted as , and the formula is:

[0040]

[0041] where includes ionization, electron excitation, vibrational excitation, and dissociation; represents the total energy - related quantity of the th energy branch (such as ionization, electron excitation, etc.); represents the set of reactions related to the th energy branch; represents the The energy corresponding to a reaction is the basic unit that constitutes . is the ionization or excitation energy, represents Avogadro's constant, represents the net chemical reaction rate of the th reaction.

[0042] The P-DRGEP method dynamically adjusts the error threshold through a closed-loop feedback mechanism to ensure that key reactions and species are retained in the simplified mechanism network, achieving an accurate match with the original mechanism network in terms of reaction dynamics and energy flow. Specifically, similar to the screening principle of the DRG algorithm, a suitable error threshold is set. If is less than the set error threshold, then the species is considered to have a very small regularized contribution to the generation rate of a certain species of the type, and the species and its related redundant reactions can be removed. If it is greater than the threshold, the species is retained. By continuously adjusting the error threshold, a skeletal mechanism with gradually decreasing numbers of species and reactions is obtained.

[0043] The depth-first search method (DFS) is an important graph search algorithm that traverses the nodes in a graph in a depth-first manner. The chemical reaction mechanism network is regarded as a graph structure, where the nodes represent reactants, intermediates, and products, and the edges represent various chemical reactions. Kinetic parameters such as reaction rates, as well as information such as the initial concentrations and temperatures of each species, are attached to the edges. The data attributes of each node and edge in the graph are constructed. Starting from the main reactant, all possible reaction paths are recursively traversed in a DFS manner.

[0044] Furthermore, starting from important starting species in the zero-carbon power system (such as or plasma-induced electronically excited states) as the search starting point, for each reaction path, its reaction flux or contribution index (such as the contribution of a certain reaction path to the generation of the final product under specific operating conditions) is calculated. During the search process, the time-step data, reaction rate constants, and transient species concentrations are used to evaluate the "weight" or importance of each reaction branch. A threshold for the reaction flux or contribution is set. If the contribution of a certain branch (i.e., a series of consecutive reactions) in the overall reaction system is lower than the set threshold, it is considered that its impact on the system response is weak. These branches are recorded during the DFS traversal and removed in subsequent steps to ensure the continuity of the core reaction chain is not disrupted (such as in the cycle path in the reaction mechanism).

[0045] Specifically, as Figure 2As shown, the depth-first search (DFS) is first used to traverse the entire skeleton mechanism to automatically identify the species containing the characteristic species (such as the cyclic reaction path ) key reaction paths; then combined with reaction path analysis (RPA), the retained reaction paths were reviewed, and the element transfer flux and energy absorption and release heat contributions in each path chain were calculated to determine the impact of each step of the reaction on rapid heating and overall energy balance.

[0046] Furthermore, the RPA reaction path analysis method is used to process the skeleton mechanism. For the retained plasma species, based on the net chemical reaction rate and chemical reaction stoichiometric coefficient of the reaction, the quantitative contribution of each path to the generation and consumption of the species is obtained, and the reaction paths with a contribution of 0 are eliminated. The search results are traversed based on DFS to ensure that the reaction paths involving the core reaction chain are not eliminated.

[0047] S3: constructing an energy network based on the key reaction path, dynamically laying out the energy network using an improved dynamic force model to obtain a dynamic energy network; optimizing the dynamic energy network using a closed-loop feedback mechanism to obtain a simplified mechanism network; The present invention is directed to a zero-carbon power system (such as In order to further improve the dynamic response and visualization level of mechanism simplification, an innovative mechanism integrating dynamic visualization analysis and intelligent feedback is proposed to further improve the dynamic response and visualization level of mechanism simplification. By deeply coupling the plasma energy transfer characteristics, reaction kinetics data and the improved dynamic force model, real-time mapping of energy contribution and closed-loop optimization of mechanism reduction are achieved.

[0048] In this embodiment, during the plasma-assisted combustion process, multiple dynamic reaction paths and the emergence and disappearance of species are involved, so the construction of a dynamic energy network is the core of achieving real-time analysis and mechanism simplification. The construction of the network is based on the following points: closed-loop screening is the distribution obtained through visualization, the error threshold of the P-DRGEP method is adjusted, and the species are used as nodes of the visualization layout. For each species: if the propagation error is small and the layout distance is large, the error threshold is reduced; otherwise, the error threshold is enlarged; all are constructed based on the data (multidimensional data) obtained by numerical simulation after preliminary reduction.

[0049] like Figure 2 As shown, S301: construct a visualization network based on the key reaction path, and stratify different species in the visualization network based on the energy type of the species to obtain an energy network.

[0050] Constructs containing plasmon active species (electron ,ion , excited state particles etc.) and key burning species ( , , A visualization network of species (e.g., , ). The attributes of each node include not only the species concentration but also its corresponding energy state (such as ionization energy, excitation energy, etc.). The edges of the visualization network represent the reaction coupling relationships between species, and the weights of the edges are determined by the reaction rate constants or the intensity of energy exchange between species.

[0051] To capture the change of species concentration over time, time-slice data is obtained by time-slicing the numerical simulation data of the reduced chemical reaction mechanism network based on the pulse period. By extracting the species concentration, energy contribution, etc. at each moment in each time slice, detailed time-series data can be obtained. The time-series data is the data of species concentration changing over time, and each time slice is combined into detailed time-series data.

[0052] By performing linear interpolation on the time-slice data, an animation of the change of species concentration over time can be generated, intuitively showing the emergence and disappearance of species. This process provides an important reference for analyzing the dynamic behavior of species during plasma discharge. Based on the energy type of species (such as ionization, excitation, etc.), different species are stratified in the generated visualization network, and species at different energy levels are distributed on the Z-axis to obtain an energy network. That is to say, through linear interpolation, the segmented time steps are made consistent, and generating an animation means visualizing each time step and then dynamically displaying it, classifying plasma species so that species at different energy levels are distributed differently, making the visualization clearer.

[0053] S302: Use an improved dynamic force model to perform dynamic layout on the energy network, so that the distribution of the energy network in two-dimensional or three-dimensional space can reflect the emergence and disappearance of species during the combustion process in real time, and a dynamic energy network is obtained.

[0054] Regarding the energy flow characteristics of plasma-assisted combustion, an energy-sensitive force mechanism is introduced into the classical ForceAtlas2 model to obtain an improved dynamic force model. It specifically includes the following three sub-modules: The energy network has been constructed previously, including node-edge weights. The repulsive force, gravitational force, and random perturbation defined below determine the distance distribution of these nodes in the energy network.

[0055] Repulsive force model: To highlight the core nodes of the energy network, the repulsive force between nodes is inversely proportional to the P-DRGEP error propagation value (or uncertainty propagation coefficient), and the error propagation value is defined as follows:

[0056] where represents the error propagation value, represents the species and the The direct interaction coefficient between energy branch variables.

[0057] The repulsive force between nodes is defined as follows:

[0058] Where, represents the repulsive force between nodes; represents the repulsive force constant; represents the minimum value, to avoid the denominator being zero; represents the distance between nodes. The above definition can ensure that high-importance species are automatically separated in the layout, forming a core node cluster.

[0059] Gravitational model: The gravitational force generated by strong reaction paths (i.e., paths with high reaction rates) will pull the relevant nodes closer together, causing them to cluster in the energy network. Therefore, the reaction path gravitational force is positively correlated with the net reaction rate, that is:

[0060] Where, represents the reaction path gravitational force, represents the gravitational constant. Strong reaction paths pull the nodes closer through gravity, forming a tight functional module.

[0061] Random perturbation term: To simulate the network layout fluctuations caused by parameter uncertainties, a random perturbation term is added and defined as:

[0062] Where, represents the random perturbation term, represents the perturbation coefficient, represents a normal distribution random variable with a mean of 0 and a standard deviation of

[0063] S303: Optimize the dynamic energy network using a closed-loop feedback mechanism to obtain a simplified mechanism network.

[0064] Intelligent feedback is achieved through a three-step cyclic iteration between the visualization results (the energy network after dynamic layout) and mechanism reduction, forming a closed loop of "layout analysis - strategy adjustment - iterative optimization". The specific process is as follows: Layout analysis: Use an improved dynamic force model to perform spatial layout on the energy network, and visually sort the species nodes according to reaction intensity and influence to obtain the node importance ranking. High-importance nodes are marked with a golden border to assist in identifying the dominant reaction chain; redundant reaction paths are marked with dotted lines to indicate candidate paths that can be removed; the network edge weights are the direct interaction coefficients calculated by P-DRGEP ​Or the error propagation coefficient definition is used to calibrate the importance of the reaction path.

[0065] Furthermore, the method for judging high-importance nodes: Calculate the rate weighted sum of a node in all participating reactions to obtain the rate contribution. Set an importance threshold (such as the top 20%) to calibrate high-importance nodes, and calculate the average distance between each pair according to the spatial coordinates of the high-importance nodes.

[0066] Furthermore, the method for judging redundant reaction paths: Edge weight screening (dashed lines are used to mark those below the threshold) and the existence of alternative paths, that is, there are multiple parallel paths between two nodes, and the weight of one of the paths is significantly lower than that of other paths, then this path can be regarded as redundant and can be represented by a dashed line for clearer visualization.

[0067] Adopt a feedback adjustment strategy for strategy adjustment: For each target species, calculate the average distance between it and the nearest several (such as 5) neighbor nodes with the strongest direct interaction in the current layout. Record the reference value (initial average distance) of these average distances in the first iteration, and use this reference value as a reference in the following. When in a certain iteration, the average distance of a certain species exceeds 120% of the "reference distance", it can be considered that this species is "too loose" in the layout, and then mark this species as "loose"; if the average distance is less than 80% of the "reference distance", it is considered that this species is "too aggregated (too close in distance)", and then mark this species as "too aggregated".

[0068] In each iteration, for the species marked as "loose", reduce its error threshold in the P-DRGEP method by 5%; for the species marked as "too aggregated", increase the error threshold by 5%. If a species belongs to neither "loose" nor "too aggregated", its error threshold can be kept unchanged, or slightly decreased (±2%) to avoid long-term stagnation. After updating the single-species threshold, evaluate the average contribution of the four major channels of ionization, electron excitation, vibrational excitation, and molecular dissociation in this round. If the overall contribution is too high (indicating that the threshold is too low and too many paths are retained), then simultaneously increase the error thresholds of these four channels by 5%; if the overall contribution is too low (excessive elimination), then simultaneously decrease by 5%. And set the lower limit to 20% of the median of all direct interaction coefficients of this species and the upper limit to 200% of the median to prevent the error threshold from being adjusted too low resulting in network distortion, or too high to eliminate key reactions.

[0069] Iterative optimization: The convergence is judged as when the average distance of a certain species is within the range of "80% - 120% reference" and its error threshold has not changed by more than 2% in three consecutive iterations, it can be considered that the threshold of this species has converged and it will no longer participate in subsequent adjustments. When all target species and the four major channels meet the above convergence conditions, the entire closed-loop can be stopped and the final skeletal mechanism (simplified mechanism network) can be output.

[0070] S4: Obtain its multi-dimensional data based on the simplified mechanism network and perform preprocessing, and perform principal component analysis on the preprocessed multi-dimensional data to obtain low-dimensional feature vectors; The numerically simulated data after order reduction (this data is obtained by further numerical simulation through the simplified mechanism network) usually contains multiple variables such as temperature and concentration. Directly using these high-dimensional data for modeling may lead to problems such as too high model complexity and feature redundancy. Therefore, the present invention first performs normalization preprocessing on each variable to ensure that they are within the same dimension range. Next, the dimensionality reduction process is performed on the normalized data through the principal component analysis (PCA) method. The specific steps include: S401: Perform normalization preprocessing on the numerically simulated data after order reduction to obtain normalized data; The present invention adopts a scaling method in the data preprocessing stage and compares common normalization techniques (such as standardization, Min-Max normalization, and Pareto method). Research shows that the Pareto normalization method can effectively scale the data while better retaining the relative differences and main variation information between variables, and avoiding the adverse effects of extreme values on the overall feature extraction. Therefore, the present invention preferably uses the Pareto normalization method to standardize the multi-variable numerical data (including temperature, concentration, etc.) generated after mechanism order reduction, so as to provide more stable and representative low-dimensional feature vectors for subsequent principal component analysis and dimensionality reduction.

[0071] S402: Perform dimensionality reduction on the normalized data through principal component analysis to obtain low-dimensional feature vectors.

[0072] Perform dimensionality reduction on the normalized data through the PCA method, calculate the principal components and load coefficients, and screen out the key variables that have a major contribution to the system response from them to form low-dimensional feature vectors.

[0073] Specifically, by solving the eigenvalue problem of the covariance matrix the main features of the data can be extracted. Let be the data matrix after normalization preprocessing, and its covariance matrix is defined as:

[0074] where, represents the number of samples, represents the feature vector matrix, represents the maximum eigenvalue, represents the transpose matrix of the feature vector matrix.

[0075] By performing By performing eigenvalue decomposition, each principal component and its corresponding importance can be obtained. Usually, the first several principal components with higher contribution rates are selected to form a low-dimensional feature vector after dimensionality reduction for subsequent machine learning modeling.

[0076] S5: Input the low-dimensional feature vector into a pre-trained comprehensive mapping model for prediction to obtain the prediction result of the key combustion parameters.

[0077] In this embodiment, the present invention proposes a multi-model fusion prediction framework. This framework uses a variety of regression techniques to construct a basic prediction model and integrates them through a non-linear combination method to form a comprehensively mapped model that has undergone strict cross-validation, which can accurately capture the multi-scale and non-linear coupling relationships in complex reaction processes and achieve real-time prediction of key combustion parameters.

[0078] The specific implementation is as follows: (1) A candidate model library including linear regression, ridge regression, decision tree, random forest, and neural network, etc. is constructed. Each model has its own advantages in terms of modeling accuracy, generalization ability, and fitting ability for non-linear relationships. In the specific prediction process, instead of simply performing equal-weight fusion on all models, the performance of candidate models is quantitatively evaluated through performance evaluation indicators, and the corresponding fusion weights are determined based on their prediction errors in the historical dataset. For samples in different working intervals or data subspaces, the present invention also supports a local adaptive strategy to dynamically select the optimal sub-model combination to improve the prediction accuracy.

[0079] Furthermore, dynamic selection is based on the average error obtained from cross-validation to evaluate the generalization ability of each candidate model on different data subspaces. The smaller the average error, the stronger the generalization ability of the model on this data subspace. The smaller the average error, the greater the weight of this model in subsequent combinations. Sub-model combination: A weight is assigned to each selected model, and the final prediction result is the weighted sum of the prediction results of each model. The weight can be determined according to the average error of cross-validation, and the model with a smaller error has a higher weight.

[0080] (2) Construct a comprehensive mapping model based on the candidate model library. To improve the generalization ability of the comprehensive mapping model and avoid overfitting problems, all candidate models use Training and performance evaluation are carried out in the way of k-fold cross-validation. The specific methods include: First, the original multi-dimensional data is dimensionally reduced to extract the principal component variables representing the system evolution characteristics; Second, the error metrics are calculated for each candidate model on the training set and the validation set respectively, and the evaluation results of all cross-validation rounds are integrated as the basis for optimizing the fusion weights; Finally, an adaptive optimization method is used to determine the contribution ratio of each candidate model in the fusion process, and a comprehensive mapping model with global prediction ability is constructed. The comprehensive model is expressed by non-linear combination as:

[0081] where, represents the final prediction result, represents the total number of candidate models, represents the th candidate model's predicted output for the input , represents the fusion weight of the th candidate model, satisfying .

[0082] (3) Input the low-dimensional feature vector into the comprehensive mapping model for prediction, and output the time series prediction results of the key combustion parameters. In the model inference stage, the system uses the preprocessed and dimensionally reduced low-dimensional feature vector as the input, which includes parameters closely related to the plasma-combustion coupling process such as species concentration and reaction temperature. Each candidate model in the comprehensive mapping model receives the same input and independently outputs the prediction results, and then the corresponding outputs are weighted according to the trained weight parameters of each candidate model to form the final predicted value. The system output includes the time series prediction results of the key combustion parameters, such as temperature evolution and main species concentration. Taking the time series prediction results as the input, driving the digital twin model of the zero-carbon power system to simulate the combustion behavior under different working conditions and accelerate the design iteration (reduce the cost of physical experiments).

[0083] (4) Determine the accuracy of the final prediction result of the comprehensive mapping model and adaptively update the model.

[0084] By dynamically monitoring the error between the final prediction result and the real data, using various metrics such as mean square error (MSE) and mean absolute error (MAE), determine the prediction accuracy of the model under the current working conditions. If the fusion model performs excellently in historical cross-validation and real-time monitoring, and the error metrics are stable within the set threshold, it is regarded as the overall optimal prediction model.

[0085] In summary, while ensuring the integrity of the key combustion reaction mechanism, the present invention achieves efficient reduction and data-driven prediction, significantly improving the efficiency and accuracy of digital modeling and real-time simulation, and providing strong technical support for the intelligent control, online monitoring, and optimal design of zero-carbon power systems.

[0086] Embodiment 2 This embodiment discloses a zero-carbon power system modeling system based on closed-loop feedback optimized reduction, including: A data acquisition module, which is configured to: acquire the chemical reaction mechanism network and its multi-dimensional data of the zero-carbon power system; A mechanism reduction module, which is configured to: perform mechanism reduction on the chemical reaction mechanism network based on the multi-dimensional data to obtain the key reaction path; A closed-loop feedback optimization module, which is configured to: construct an energy network based on the key reaction path, perform dynamic layout on the energy network using an improved dynamic force model to obtain a dynamic energy network; optimize the dynamic energy network using a closed-loop feedback mechanism to obtain a simplified mechanism network; A feature extraction module, which is configured to: acquire the multi-dimensional data of the simplified mechanism network and perform preprocessing, and perform principal component analysis on the preprocessed multi-dimensional data to obtain a low-dimensional feature vector; A model prediction module, which is configured to: input the low-dimensional feature vector into a pre-trained comprehensive mapping model for prediction to obtain the prediction result of the key combustion parameters.

[0087] Embodiment 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the method in Embodiment 1 are implemented.

[0088] Embodiment 4 The purpose of this embodiment is to provide a computer-readable storage medium. A computer-readable storage medium stores a computer program, and when the program is executed by a processor, the steps of the method in Embodiment 1 are executed.

[0089] The steps involved in the devices in Embodiments 3 and 4 above correspond to those in Method Embodiment 1. For specific implementation manners, reference may be made to the relevant description part of Embodiment 1. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0090] Those skilled in the art should understand that the various modules or steps of the present invention described above can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0091] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0092] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications or deformations that can be made without creative efforts by those skilled in the art on the basis of the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction, characterized in that: include: Obtain the chemical reaction mechanism network and its multi-dimensional data of zero-carbon power systems; Based on the multidimensional data, the chemical reaction mechanism network is reduced to obtain a key reaction path; An energy network is constructed based on the key reaction path, and an improved dynamic force model is used to dynamically layout the energy network to obtain a dynamic energy network; The dynamic energy network is optimized by adopting a closed-loop feedback mechanism to obtain a simplified mechanism network; Based on the simplified mechanism network, the multidimensional data is obtained and preprocessed, and the principal component analysis is performed on the preprocessed multidimensional data to obtain a low-dimensional feature vector; The low-dimensional feature vector is input into a pre-trained comprehensive mapping model for prediction to obtain prediction results of key combustion parameters.

2. The zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as claimed in claim 1 is characterized in that: The chemical reaction mechanism network is reduced to a direct relationship graph method with plasma error propagation to calculate the direct interaction coefficient between species and each energy branch variable. When the direct interaction coefficient is less than the set error threshold, the corresponding species and their related redundant reactions are eliminated to obtain the skeleton mechanism.

3. The zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as claimed in claim 2 is characterized in that: First, the skeleton mechanism is traversed using depth-first search to automatically identify key reaction paths, and then the retained reaction paths are reviewed in conjunction with reaction path analysis.

4. The zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as claimed in claim 1 is characterized in that: Constructing an energy network based on key reaction paths is specifically as follows: constructing a visualization network based on the key reaction paths, and stratifying different species in the visualization network based on the energy types of the species to obtain an energy network.

5. The zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as claimed in claim 1, characterized in that: An improved dynamic force model is obtained by introducing an energy-sensitive force mechanism into the ForceAtlas2 model, including a repulsive force model, a gravitational force model and a random disturbance term.

6. The zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as claimed in claim 1, characterized in that: The closed-loop feedback mechanism is specifically: The improved dynamic force model is used to spatially layout the energy network, and the species nodes are visually sorted according to the reaction intensity and influence to obtain the node importance ranking; Based on the node importance ranking, high-importance nodes are marked, and the average distance between each of the high-importance nodes is calculated according to the spatial coordinates of the high-importance nodes; Based on the average distance, a feedback adjustment strategy is used to adjust the error threshold in the directed relationship graph method of plasma band error propagation, and the mechanism is reduced based on the adjusted error threshold; The above steps are iterated. When the error thresholds of all species nodes meet the convergence conditions, the closed loop is stopped and the simplified mechanism network is output.

7. The zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as claimed in claim 1, characterized in that: A comprehensive mapping model is formed based on a multi-model fusion prediction framework, specifically: Building a candidate model library containing multiple candidate models; All candidate models are adopted The training and performance evaluation are carried out by means of fold cross validation, and the fusion weight of each candidate model is obtained based on the performance evaluation; A comprehensive mapping model is constructed based on each candidate model and the corresponding fusion weights.

8. A zero-carbon power system modeling system based on closed-loop feedback optimization and order reduction, characterized in that: include: A data acquisition module, which is configured to: acquire a chemical reaction mechanism network of a zero-carbon power system and its multi-dimensional data; A mechanism reduction module is configured to: perform mechanism reduction on the chemical reaction mechanism network based on the multidimensional data to obtain a key reaction path; A closed-loop feedback optimization module is configured to: construct an energy network based on the key reaction path, dynamically layout the energy network using an improved dynamic force model to obtain a dynamic energy network; optimize the dynamic energy network using a closed-loop feedback mechanism to obtain a simplified mechanism network; A feature extraction module is configured to: obtain and preprocess the multidimensional data based on the simplified mechanism network, and perform principal component analysis on the preprocessed multidimensional data to obtain a low-dimensional feature vector; The model prediction module is configured to: input the low-dimensional feature vector into a pre-trained comprehensive mapping model for prediction, and obtain prediction results of key combustion parameters.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the zero-carbon power system modeling method based on closed-loop feedback optimization and order reduction as described in any one of claims 1 to 7 are implemented.

Images