Biomass pyrolysis nitrogen prediction and question and answer system based on machine learning and agent

Abstract

This invention provides a biomass pyrolysis nitrogen prediction and question-answering system based on machine learning and intelligent agents. This invention incorporates machine learning feature engineering methods to screen various features for predicting the distribution of pyrolysis nitrogen components, extracting the core features most important to the distribution, thus reducing numerous parameters and repetitive experimental work. By constructing a pyrolysis nitrogen component distribution prediction model, it consumes less manpower, financial resources, and time, resulting in lower costs and a shorter cycle. Furthermore, traditional pyrolysis nitrogen prediction methods typically only predict without interaction, making it impossible to effectively verify the prediction results. The question-answering system of this invention not only enables interaction between the user and the system but also uses a combination of large language models and dedicated intelligent agent invocation to match users with more accurate answers.

Description

Technical Field

[0001] This invention belongs to the field of nitrogen prediction technology, and relates to a biomass pyrolysis nitrogen prediction and question-answering system based on machine learning and intelligent agents. Background Technology

[0002] As the only carbon-based renewable resource, biomass energy plays a crucial role in achieving my country's strategic goal of "carbon peaking and carbon neutrality" through its high-value utilization. Pyrolysis technology, as a promising low-carbon conversion method, can transform biomass into high-value-added energy and chemical products such as pyrolysis gas, bio-oil, and biochar, and is one of the key pathways to replace fossil fuels and alleviate the energy crisis and environmental pressure.

[0003] Biomass is diverse, with nitrogen content ranging from 0.1% to 9 wt.%. During pyrolysis, nitrogen from the feedstock is distributed in different forms into the gaseous, liquid, and solid three-phase products, affecting the quality and subsequent utilization of the pyrolysis products. For example, nitrogen in pyrolysis gas and bio-oil is the main source of NOx during their subsequent combustion, which is detrimental to their fuel properties. On the other hand, nitrogen-containing functional groups in biochar can significantly improve its surface activity and enhance its adsorption, catalytic, and electrochemical properties. Therefore, understanding the distribution pattern of nitrogen from the feedstock into the three-phase products during biomass pyrolysis is fundamental to controlling the quality of pyrolysis products and is a key scientific issue for achieving high-value and clean utilization of biomass, which has always been a research focus in the biomass field.

[0004] Traditional methods require numerous parameters to predict nitrogen in pyrolysis, and the migration and transformation of nitrogen in biomass pyrolysis relies on a large number of repetitive experiments. Characterization tests consume a lot of manpower, financial resources, and time, resulting in high costs and long cycles. Summary of the Invention

[0005] To address the problems of high cost and long cycle associated with the current reliance on numerous repetitive experiments for nitrogen migration and transformation in biomass pyrolysis, and the significant human, financial, and time-consuming characterization and testing of pyrolysis products, this invention provides a biomass pyrolysis nitrogen prediction and question-answering system based on machine learning and intelligent agents. This invention incorporates machine learning feature engineering methods to screen various features for predicting the distribution of pyrolysis nitrogen components, extracting the core features most important to the distribution, thus reducing numerous parameters and repetitive experimental work. By constructing a pyrolysis nitrogen component distribution prediction model, it consumes less human, financial, and time, resulting in lower costs and a shorter cycle. Furthermore, traditional pyrolysis nitrogen prediction methods typically only predict without interaction, failing to effectively verify the prediction results. This invention's question-answering system not only enables interaction between the user and the system but also uses a combination of large language models and dedicated intelligent agent invocation to provide users with more accurate answers.

[0006] The specific technical solution of the present invention is as follows:

[0007] A biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents includes:

[0008] The data preprocessing module is used to collect biomass raw material attribute data, pyrolysis process parameter data and measured pyrolysis nitrogen release data, and perform data cleaning, missing value filling and standardization to obtain a standardized dataset.

[0009] The feature engineering module performs feature extraction, feature filtering, and feature transformation on the standardized dataset to generate a pyrolysis nitrogen prediction feature set. The features include biomass elemental composition features, raw material physical property features, and pyrolysis process features.

[0010] The intelligent prediction module uses a machine learning model built on the CatBoost algorithm to predict the nitrogen content and nitrogen-containing component distribution of the three-phase products based on the key feature set selected by feature engineering.

[0011] As a preferred embodiment of the present invention, the feature engineering module includes:

[0012] Filtering selection includes the Pearson correlation coefficient (PCC) method, which uses the Pearson correlation coefficient to quickly remove strong linear relationships between features; the mutual information (MI) method, which captures non-linear relationships between features and targets; and the variance thresholding (VT) method, which reduces dimensionality by removing low-variance features and selecting features with large variance.

[0013] Embedded selection includes Random Forest (RF) which uses a tree model to compute the importance ranking of each feature for the prediction result, and Minimal Absolute Shrinkage and Selection Operator (LASSO) which compresses the coefficients to 0 by applying an L1 penalty term to the coefficients.

[0014] Feature transformation methods include principal component analysis (PCA) which converts the original features into a small number of low-dimensional features by constructing a covariance matrix; partial least squares (PLS) which considers the multivariate covariance between features and between features and the target; autoencoders (AE) which convert input data into low-dimensional representations and input them into the latent layer; and variational autoencoders (VAE) which compress the input data through feature extraction capabilities.

[0015] As a preferred embodiment of the present invention, the intelligent prediction module includes the following prediction method:

[0016] S1. Data input and preprocessing: The collected biomass raw material attribute data, pyrolysis process parameter data, and pyrolysis nitrogen release measured data are cleaned, missing value filled, and standardized to obtain a standardized dataset.

[0017] S2. Feature Extraction: Using filtering methods such as Pearson correlation coefficient (PCC), mutual information (MI), and variance thresholding (VT), and embedded selection methods such as random forest (RF) and least absolute shrinkage and selection operator (LASSO), and feature transformation methods such as principal component analysis (PCA), partial least squares (PLS), autoencoder (AE), and variational autoencoder (VAE), features are extracted, filtered, and transformed from the standardized dataset to select core features and generate a pyrolysis nitrogen prediction feature set. These features include biomass elemental composition characteristics, raw material physical property characteristics, and pyrolysis process characteristics.

[0018] S3. Model Construction: A machine learning prediction model was constructed for each prediction target. Considering the characteristics of the dataset, such as small sample size, non-linear relationship, and classification features, the CatBoost algorithm was selected, and 13×9=117 machine learning prediction models based on feature engineering were constructed. Cross-validation and other methods were used to train and optimize the hyperparameters of the above models, and the root mean square error (RMSE) and coefficient of determination (R²) were used as the main evaluation indicators of model performance.

[0019] S4, Agent Invocation: Connect the model analysis results to the agents, and use a multi-agent collaborative architecture to invoke three agents to provide accurate answers to the user.

[0020] As a preferred embodiment of the present invention, the intelligent prediction module includes a model optimization unit, which uses a cross-validation algorithm to optimize the hyperparameters of the CatBoost model in order to minimize the root mean square error (RMSE) between the prediction results and the measured data and maximize the coefficient of determination (R²).

[0021] As a preferred embodiment of the present invention, the prediction target of the intelligent prediction module includes the nitrogen content of the three-phase products and the specific distribution of nitrogen-containing components, specifically including one or more combinations of the following: pyridine nitrogen N-5, pyrrole nitrogen N-6, quaternary nitrogen NQ, amino nitrogen NA, other nitrogen-containing compounds NX, amine nitrogen N-ami, aliphatic / aromatic nitrogen-containing compounds N-com, heterocyclic nitrogen-containing compounds N-het, nitro nitrogen N-nit, NH3 and HCN.

[0022] As a preferred embodiment of the present invention, the multi-agent question-answering module supports incremental learning, allowing for real-time updates of data, rules, and dialogue strategies of the agents in the knowledge base.

[0023] As a preferred embodiment of the present invention, the system is a desktop application software with an interface divided into an independent parameter input prediction area and a natural language question-and-answer interaction area, supporting users to obtain real-time predictions and decision suggestions through a graphical interface.

[0024] The biomass pyrolysis nitrogen question-answering system based on machine learning and intelligent agents, and the biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents mentioned above, further include:

[0025] The multi-agent question-answering module adopts a multi-agent collaborative architecture. It uses a central agent to allocate tasks and automatically match professional agents. Combined with search enhancement generation technology, it provides users with question-answering services in the field of biomass pyrolysis.

[0026] The data interaction module provides a display of prediction results and a natural language question-and-answer interface;

[0027] The multi-agent question-answering module includes specialized agents such as biomass pyrolysis agents, biochar agents, and nitrogen-related agents, each of which is associated with a structured professional knowledge base.

[0028] As a preferred embodiment of the present invention, the multi-agent question answering module is implemented through a retrieval enhancement generation unit: the user's question is semantically similar to the internal knowledge base, relevant knowledge fragments are retrieved as context, and submitted to the large language model along with the user's question to generate an answer.

[0029] As a preferred embodiment of the present invention, the construction of the knowledge base involves using OCR technology to process academic papers and technical reports in the field, and employing ontology construction technology to establish a conceptual system, attribute relationships, and rule base in the field of biomass pyrolysis nitrogen.

[0030] The beneficial effects of this invention are as follows:

[0031] 1. This invention uses the CatBoost algorithm to build a machine learning prediction model based on feature engineering. For each prediction target, a machine learning prediction model is built, and the core important features that affect each target are selected using feature engineering methods. The prediction results are based on these selected core features, and the results are more consistent with reality.

[0032] 2. This invention uses a multi-agent collaborative architecture, which enables three specialized agents—biomass pyrolysis agent, biochar agent, and nitrogen-related agent—to perform their respective functions while interacting with each other. A central agent is used to allocate tasks and automatically match the specialized agents that are relevant to the user's input, providing the user with accurate answers.

[0033] 3. This invention uses a dedicated knowledge base to systematically collect and organize authoritative textbooks, high-quality academic papers, and technical reports in the field of content cleaning. It enables OCR technology in the content processing rules and configures index enhancement strategies to build a structured professional knowledge base for each agent. When a user inputs a question, the system first performs semantic similarity matching between the question and the internal knowledge base, quickly retrieving the most relevant knowledge fragments. These precise professional knowledge fragments are then used as context and submitted to the large language model along with the user's question. The large language model then processes and integrates these fragments, resulting in a professional, accurate, and easily understandable answer.

[0034] 4. This invention comprises two parts: a prediction area and a question-and-answer area. After relevant parameters are input into the prediction area, the model returns the prediction results for the nitrogen content and nitrogen-containing component distribution of the gas, liquid, and solid three-phase products. The question-and-answer area provides an intelligent question-and-answer interface. Users only need to input biomass pyrolysis-related questions in natural language, and the system can capture keywords and automatically match the corresponding intelligent agent to answer the questions, making the operation process convenient and simple. Attached Figure Description

[0035] Figure 1 This is the software core module and technology roadmap of the biomass pyrolysis nitrogen prediction and question answering system based on machine learning and intelligent agents of the present invention.

[0036] Figure 2 This is an interactive interface diagram of an embodiment of the biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents of the present invention;

[0037] Figure 3 This is an interactive interface diagram of an embodiment of the biomass pyrolysis nitrogen question-and-answer system of the present invention, which integrates machine learning and intelligent agents. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 , Figure 2 , Figure 3 A biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents includes:

[0040] The data preprocessing module is used to collect biomass raw material attribute data, pyrolysis process parameter data and measured pyrolysis nitrogen release data, and perform data cleaning, missing value filling and standardization to obtain a standardized dataset.

[0041] The feature engineering module performs feature extraction, feature filtering, and feature transformation on the standardized dataset to generate a pyrolysis nitrogen prediction feature set. The features include biomass elemental composition features, raw material physical property features, and pyrolysis process features.

[0042] The intelligent prediction module uses a machine learning model built on the CatBoost algorithm to predict the nitrogen content and nitrogen-containing component distribution of the three-phase products based on the key feature set selected by feature engineering.

[0043] This module is the core of the software's quantitative analysis. Its goal is to establish a high-precision prediction model that can accurately predict the nitrogen content and distribution of major nitrogen-containing components in the gas, liquid, and solid three-phase products based on the characteristics of the input biomass feedstock and pyrolysis process conditions.

[0044] The feature engineering module includes:

[0045] Filtering selection includes the Pearson correlation coefficient (PCC) method, which uses the Pearson correlation coefficient to quickly remove strong linear relationships between features; the mutual information (MI) method, which captures non-linear relationships between features and targets; and the variance thresholding (VT) method, which reduces dimensionality by removing low-variance features and selecting features with large variance.

[0046] Embedded selection includes Random Forest (RF) which uses a tree model to compute the importance ranking of each feature for the prediction result, and Minimal Absolute Shrinkage and Selection Operator (LASSO) which compresses the coefficients to 0 by applying an L1 penalty term to the coefficients.

[0047] Feature transformation methods include principal component analysis (PCA) which converts the original features into a small number of low-dimensional features by constructing a covariance matrix; partial least squares (PLS) which considers the multivariate covariance between features and between features and the target; autoencoders (AE) which convert input data into low-dimensional representations and input them into the latent layer; and variational autoencoders (VAE) which compress the input data through feature extraction capabilities.

[0048] Filtering-based selection can quickly eliminate redundant and irrelevant features in the original dataset at the initial stage of the feature selection process, without the need for complex model iteration and optimization, significantly reducing data dimensionality and computational overhead. Compared to full feature modeling, it effectively reduces the time and space complexity of subsequent model training, making it suitable for preprocessing scenarios with large-scale, high-dimensional datasets. By filtering out features with low relevance to the target task, it avoids the model learning noise information in the data, reducing the risk of overfitting. This technique is based on the statistical properties of features and labels for selection, independent of specific models, and has good versatility and robustness, improving the predictive stability of different types of machine learning models on new samples.

[0049] Embedded selection integrates the feature selection process into the model training process. By introducing a regularization term into the model's objective function, the model automatically filters redundant features while learning parameters, eliminating the need for a separate feature selection process. Compared to filtering and wrapping techniques, this reduces process steps, improves the adaptability of feature selection to the model, and balances feature interaction with computational efficiency. Furthermore, after filtering using this method, the model retains only features that significantly contribute to the prediction results. The sparsity of model parameters can intuitively reflect the importance of each feature. Compared to full-feature models, this simplifies the model structure, reduces the difficulty of model interpretation, and facilitates the analysis of the intrinsic relationship between features and the prediction target by technical personnel.

[0050] Feature transformation methods use mapping functions to transform nonlinearly separable data in the original low-dimensional space to a high-dimensional feature space, giving the transformed data linearly separable properties. This expands the applicability of linear classification and regression models, enabling effective modeling of nonlinear data without constructing complex nonlinear models. It also reduces redundancy and correlation between features, enhancing the robustness and expressive power of the data.

[0051] The intelligent prediction module includes the following prediction methods:

[0052] S1. Data input and preprocessing: The collected biomass raw material attribute data, pyrolysis process parameter data, and pyrolysis nitrogen release measured data are cleaned, missing value filled, and standardized to obtain a standardized dataset.

[0053] S2. Feature Extraction: Using filtering methods such as Pearson correlation coefficient (PCC), mutual information (MI), and variance thresholding (VT), and embedded selection methods such as random forest (RF) and least absolute shrinkage and selection operator (LASSO), and feature transformation methods such as principal component analysis (PCA), partial least squares (PLS), autoencoder (AE), and variational autoencoder (VAE), features are extracted, filtered, and transformed from the standardized dataset. Core features are selected to generate a pyrolysis nitrogen prediction feature set, including biomass elemental composition, raw material physical properties, and pyrolysis process characteristics.

[0054] S3. Model Construction: A machine learning prediction model was constructed for each prediction target. Considering the characteristics of the dataset, such as small sample size, non-linear relationship, and classification features, the CatBoost algorithm was selected, and 13×9=117 machine learning prediction models based on feature engineering were constructed. Cross-validation and other methods were used to train and optimize the hyperparameters of the above models, and the root mean square error (RMSE) and coefficient of determination (R²) were used as the main evaluation indicators of model performance.

[0055] S4, Agent Invocation: Connect the model analysis results to the agents, and use a multi-agent collaborative architecture to invoke three agents to provide accurate answers to the user.

[0056] The intelligent agents are:

[0057] The biomass pyrolysis intelligent agent focuses on answering general questions such as the basic principles of pyrolysis, pyrolysis technology, the influence of process parameters, and the application of products.

[0058] Biochar smart bodies focus on the preparation process, properties (such as specific surface area, functional groups, adsorption performance) and applications of biochar.

[0059] Nitrogen-related intelligent agents are specifically designed to answer professional questions regarding the migration pathways, transformation mechanisms, nitrogen-containing products, and nitrogen regulation strategies of nitrogen elements in biomass during pyrolysis.

[0060] The intelligent prediction module includes a model optimization unit that uses a cross-validation algorithm to optimize the hyperparameters of the CatBoost model in order to minimize the root mean square error (RMSE) between the prediction results and the measured data and maximize the coefficient of determination (R²).

[0061] Test results show that the feature-engineered CatBoost model performs well in predicting the nitrogen content and nitrogen-containing component distribution of three-phase products, with extremely high prediction accuracy (R2>0.9).

[0062] The intelligent prediction module predicts the nitrogen content of the three-phase products and the specific distribution of nitrogen-containing components, including one or more of the following: pyridine nitrogen N-5, pyrrole nitrogen N-6, quaternary nitrogen NQ, amino nitrogen NA, other nitrogen-containing compounds NX, amine nitrogen N-ami, aliphatic / aromatic nitrogen-containing compounds N-com, heterocyclic nitrogen-containing compounds N-het, nitro nitrogen N-nit, NH3, and HCN.

[0063] The multi-agent question-answering module supports incremental learning, allowing for real-time updates of data, rules, and the agents' dialogue strategies in the knowledge base.

[0064] We systematically collected, organized, and cleaned authoritative textbooks, high-quality academic papers, and technical reports in the field. We enabled OCR technology in the content processing rules, configured index enhancement strategies, and built a structured professional knowledge base for each agent.

[0065] The system is a desktop application software with an interface divided into an independent parameter input and prediction area and a natural language question and answer interaction area, allowing users to obtain real-time predictions and decision suggestions through a graphical interface.

[0066] When a user enters a question, the system first performs semantic similarity matching between the question and its internal knowledge base, quickly retrieving the most relevant knowledge fragments. This precise professional knowledge, used as context, is submitted along with the user's question to the large language model. The large language model then processes and integrates this information to generate a professional, accurate, and easy-to-understand answer. This design enables the agent not only to answer factual questions but also to perform a certain degree of summarization, analysis, and reasoning.

[0067] like Figure 1 , Figure 2 , Figure 3 Based on the aforementioned biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents, it also includes:

[0068] The multi-agent question-answering module adopts a multi-agent collaborative architecture. It uses a central agent to allocate tasks and automatically match professional agents. Combined with search enhancement generation technology, it provides users with question-answering services in the field of biomass pyrolysis.

[0069] The data interaction module provides a display of prediction results and a natural language question-and-answer interface;

[0070] The multi-agent question-answering module includes specialized agents such as biomass pyrolysis agents, biochar agents, and nitrogen-related agents, each of which is associated with a structured professional knowledge base.

[0071] This knowledge base employs Retrieval-Augmented Generation (RAG) technology to achieve deep binding between large language models and structured domain knowledge, significantly improving the professionalism and accuracy of the output content.

[0072] The multi-agent question answering module achieves this through a retrieval-enhanced generation unit: it performs semantic similarity matching between the user's question and the internal knowledge base, retrieves relevant knowledge fragments as context, and submits them along with the user's question to the large language model to generate the answer.

[0073] The construction of the knowledge base involves using OCR technology to process academic papers and technical reports in the field, and adopting ontology construction technology to establish a conceptual system, attribute relationships and rule base in the field of biomass pyrolysis nitrogen.

[0074] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention should fall within the protection scope of the present invention. All technical contents for which protection is sought in this invention have been fully described in the claims.

Claims

1. A biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents, characterized in that, include: The data preprocessing module is used to collect biomass raw material attribute data, pyrolysis process parameter data and measured pyrolysis nitrogen release data, and perform data cleaning, missing value filling and standardization to obtain a standardized dataset. The feature engineering module performs feature extraction, feature filtering, and feature transformation on the standardized dataset to generate a pyrolysis nitrogen prediction feature set. The features include biomass elemental composition features, raw material physical property features, and pyrolysis process features. The intelligent prediction module uses a machine learning model built on the CatBoost algorithm to predict the nitrogen content and nitrogen-containing component distribution of the three-phase products based on the key feature set selected by feature engineering.

2. The biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents according to claim 1, characterized in that, The feature engineering module includes: Filtering selection includes the Pearson correlation coefficient (PCC) method, which uses the Pearson correlation coefficient to quickly remove strong linear relationships between features; the mutual information (MI) method, which captures non-linear relationships between features and targets; and the variance thresholding (VT) method, which reduces dimensionality by removing low-variance features and selecting features with large variance. Embedded selection includes Random Forest (RF) which uses a tree model to compute the importance ranking of each feature for the prediction result, and Minimal Absolute Shrinkage and Selection Operator (LASSO) which compresses the coefficients to 0 by applying an L1 penalty term to the coefficients. Feature transformation methods include principal component analysis (PCA) which converts the original features into a small number of low-dimensional features by constructing a covariance matrix; partial least squares (PLS) which considers the multivariate covariance between features and between features and the target; autoencoders (AE) which convert input data into low-dimensional representations and input them into the latent layer; and variational autoencoders (VAE) which compress the input data through feature extraction capabilities.

3. The biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents according to claim 1, characterized in that, The intelligent prediction module includes the following prediction methods: S1. Data input and preprocessing: The collected biomass raw material attribute data, pyrolysis process parameter data, and pyrolysis nitrogen release measured data are cleaned, missing value filled, and standardized to obtain a standardized dataset. S2. Feature Extraction: Using filtering methods such as Pearson correlation coefficient (PCC), mutual information (MI), and variance thresholding (VT), and embedded selection methods such as random forest (RF) and least absolute shrinkage and selection operator (LASSO), and feature transformation methods such as principal component analysis (PCA), partial least squares (PLS), autoencoder (AE), and variational autoencoder (VAE), features are extracted, filtered, and transformed from the standardized dataset to select core features and generate a pyrolysis nitrogen prediction feature set. These features include biomass elemental composition characteristics, raw material physical property characteristics, and pyrolysis process characteristics. S3. Model Construction: A machine learning prediction model was constructed for each prediction target. Considering the characteristics of the dataset, such as small sample size, non-linear relationship, and classification features, the CatBoost algorithm was selected, and 13×9=117 machine learning prediction models based on feature engineering were constructed. Cross-validation and other methods were used to train and optimize the hyperparameters of the above models, and the root mean square error (RMSE) and coefficient of determination (R²) were used as the main evaluation indicators of model performance. S4, Agent Invocation: Connect the model analysis results to the agents, and use a multi-agent collaborative architecture to invoke three agents to provide accurate answers to the user.

4. The biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents according to claim 1, characterized in that, The intelligent prediction module includes a model optimization unit that uses a cross-validation algorithm to optimize the hyperparameters of the CatBoost model in order to minimize the root mean square error (RMSE) between the prediction results and the measured data and maximize the coefficient of determination (R²).

5. The biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents according to claim 1, characterized in that, The prediction targets of the intelligent prediction module include the nitrogen content of the three-phase products and the specific distribution of nitrogen-containing components, specifically including one or more combinations of the following: pyridine nitrogen N-5, pyrrole nitrogen N-6, quaternary nitrogen NQ, amino nitrogen NA, other nitrogen-containing compounds NX, amine nitrogen N-ami, aliphatic / aromatic nitrogen-containing compounds N-com, heterocyclic nitrogen-containing compounds N-het, nitro nitrogen N-nit, NH3 and HCN.

6. The biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents according to claim 1, characterized in that, The multi-agent question-answering module supports incremental learning, allowing for real-time updates of data, rules, and the agents' dialogue strategies in the knowledge base.

7. The biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents according to claim 1, characterized in that, The system is a desktop application software with an interface divided into an independent parameter input prediction area and a natural language question-and-answer interaction area, allowing users to obtain real-time predictions and decision suggestions through a graphical interface.

8. A biomass pyrolysis nitrogen question-and-answer system based on machine learning and intelligent agents, characterized in that, The biomass pyrolysis nitrogen prediction system based on machine learning and intelligent agents according to any one of claims 1-7 further includes: The multi-agent question-answering module adopts a multi-agent collaborative architecture. It uses a central agent to allocate tasks and automatically match professional agents. Combined with search enhancement generation technology, it provides users with question-answering services in the field of biomass pyrolysis. The data interaction module provides a display of prediction results and a natural language question-and-answer interface; The multi-agent question-answering module includes specialized agents such as biomass pyrolysis agents, biochar agents, and nitrogen-related agents, each of which is associated with a structured professional knowledge base.

9. The biomass pyrolysis nitrogen question-and-answer system based on machine learning and intelligent agents according to claim 8, characterized in that, The multi-agent question answering module is implemented through a retrieval enhancement generation unit: it performs semantic similarity matching between the user's question and the internal knowledge base, retrieves relevant knowledge fragments as context, and submits them along with the user's question to the large language model to generate the answer.

10. The biomass pyrolysis nitrogen question-and-answer system based on machine learning and intelligent agents according to claim 9, characterized in that, The construction of the knowledge base involves using OCR technology to process academic papers and technical reports in the field, and employing ontology construction technology to establish a conceptual system, attribute relationships, and rule base for the field of biomass pyrolysis nitrogen.

Images