Human-land interaction simulation method and system coupling large language model and process model
Patent Information
- Application Number
- CN202610971755.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-01
- Publication Date
- 2026-10-09
AI Technical Summary
[0002]现有人类—环境耦合模拟通常采用生物地球化学模型、生态过程模型或其他自然过程模型对环境指标变化进行分析，但在实际应用中，人类管理行为常被简化为固定参数或预设情景，难以反映政策、价格、补贴、培训等因素变化下的异质性行为响应，进而无法准确农田氮循环中人类活动与自然过程相互租用的动态预测
[0029]1. 本发明将农户行为预测与农业生物地球化学过程模拟进行耦合，能够实现从政策激励、农户施肥响应到氮排放结果的链式量化分析，克服了现有技术中仅关注单一环节、难以反映政策作用全过程的问题。
Smart Images
Figure CN122886356A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of farmland ecosystem simulation and agricultural environmental policy assessment, specifically involving a human-land interaction simulation method and system that couples a large language model and a process model. Background Technology
[0002] Existing human-environment coupling simulations typically use biogeochemical models, ecological process models, or other natural process models to analyze changes in environmental indicators. However, in practical applications, human management behavior is often simplified to fixed parameters or preset scenarios, making it difficult to reflect heterogeneous behavioral responses under changes in factors such as policies, prices, subsidies, and training. Consequently, it is impossible to accurately predict the dynamic inter-rental interaction between human activities and natural processes in the nitrogen cycle of farmland. Summary of the Invention
[0003] The purpose of this invention is to provide a method and system for simulating human-environment interaction by coupling a large language model and a process model, so as to reduce the bias caused by static management scenario assumptions and improve the dynamic representation ability of human management behavior response and its environmental effects under policy scenarios.
[0004] To achieve the above objectives, the present invention proposes the following technical solution:
[0005] A method for simulating human-land interaction by coupling a large language model and a process model includes the following steps:
[0006] Based on survey and statistical data of farmers in the target area, a standardized sample set is constructed for training a large language model;
[0007] A pre-trained large language model is selected, and the parameters of the large language model are fine-tuned using the standardized sample set to obtain a farmer decision-making agent model that can generate inference results on changes in fertilization behavior based on the input farmer attributes and policy scenarios.
[0008] During the reasoning phase, the attributes and preset policy variables of each farmer in the target area are input into the farmer decision-making agent model to obtain the fertilization adjustment results corresponding to each farmer.
[0009] The fertilization adjustment results are parameterized and mapped to management input parameters that can be recognized by the gridded biogeochemical model. The parameterization process includes at least: scaling the baseline fertilization amount according to the reduction ratio in the fertilization adjustment results, and establishing a correspondence between the scaled fertilization amount and spatial grid units or plot units.
[0010] The gridded biogeochemical model is driven to run in parallel within the target area, in conjunction with the management input parameters, to simulate the process of environmental indicator changes.
[0011] Preferred methods for constructing standardized sample sets for training large language models include:
[0012] The input content is formed by combining farmers' characteristic information with policy context information, and the output content is formed by the corresponding changes in fertilization behavior. Training samples in the format of "instruction-input-output" are constructed according to the preset instruction template.
[0013] Preferably, constructing the standardized sample set further includes:
[0014] Interpolation is performed on samples corresponding to adjacent policy scenarios, and perturbations are added to the output results to expand the sample size and improve the model's generalization ability to unobserved policy scenarios.
[0015] Preferably, when fine-tuning the parameters of the pre-trained large language model, the target hyperparameter combination for model fine-tuning is determined by searching multiple training hyperparameter combinations.
[0016] Preferably, the process of determining the combination of the multiple training hyperparameters includes:
[0017] Different training rounds, batch sizes, gradient accumulation steps, and learning rates are set, and the models are evaluated based on repeatedly partitioned training and test sets to select farmer decision-making agent models that meet the preset performance conditions.
[0018] Preferably, after obtaining the fertilizer adjustment results for each farmer, the process also includes:
[0019] The fertilization adjustment results are allocated to spatial grid cells or plot cells according to preset rules to form a set of management input parameters corresponding to the gridded biogeochemical model.
[0020] Preferably, the environmental indicators include NH3 volatilization, non-point source nitrogen (NPSN), N2O and CH4 emissions.
[0021] This invention also proposes a human-land interaction simulation system that couples a large language model and a process model, comprising:
[0022] The data preprocessing module is used to acquire survey data and statistical data of farmers in the target area, and to standardize the data to construct a sample set for training a large language model; the data preprocessing module is also used to organize farmer characteristic information and policy scenario information into a preset input format, and to organize the corresponding fertilization behavior change results into an output format, so as to form standardized data that can be used for model training and inference.
[0023] The farmer decision-making agent module is used to call the pre-trained large language model and perform efficient parameter fine-tuning of the large language model based on the standardized sample set to construct the farmer decision-making agent model; the farmer decision-making agent module is also used to receive farmer attribute information and preset policy variables during the inference stage and output the corresponding fertilization adjustment results of the farmers under the policy scenario.
[0024] The coupling interface module is used to unify and parameterize the fertilization adjustment results and map them into management input parameters that can be recognized by the gridded biogeochemical model. The coupling interface module is also used to scale the baseline fertilization amount according to the fertilization adjustment results and establish a correspondence between the scaled management input parameters and spatial grid units or plot units.
[0025] The gridded biogeochemical model simulation module is used to receive the management input parameters and drive the gridded biogeochemical model to perform grid-by-grid parallel simulations in the target area to obtain the results of environmental indicator changes.
[0026] The results output and analysis module is used to summarize, statistically analyze, and spatially visualize the results of changes in the environmental indicators, and generate the marginal emission reduction and pollution reduction effects and their spatial distribution results under different policy scenarios.
[0027] Preferably, the environmental indicator changes include NH3 volatilization, non-point source nitrogen (NPSN), N2O emissions, and CH4 emissions.
[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0029] 1. This invention couples farmer behavior prediction with agricultural biogeochemical process simulation, enabling chain-like quantitative analysis from policy incentives and farmer fertilization responses to nitrogen emission results. This overcomes the problem in existing technologies that only focus on a single link and are difficult to reflect the entire process of policy effects.
[0030] 2. This invention uses a fine-tuned large language model to reason about farmers' fertilization decisions. It can predict farmers' response trends under different subsidy or price scenarios even in the absence of large-scale policy implementation observation data or long-term pilot data. This provides a technical approach for agricultural environmental policies that has lower data requirements, lower implementation costs, and is suitable for ex-ante assessment.
[0031] 3. This invention can further map changes in farmers' fertilization behavior into environmental effects such as NH3 volatilization, N2O emissions, CH4 emissions, and non-point source nitrogen loss, thereby achieving a quantitative assessment of the environmental consequences of policies and improving the completeness and practical value of policy analysis results.
[0032] 4. This invention has the ability to identify spatial heterogeneity and update dynamically, which can support the design of differentiated policies and the optimization of subsidy intensity in different regions, and can continuously correct the prediction results by combining real-time observation data during the policy implementation process, thereby improving the scientificity and accuracy of policy formulation and optimization. Attached Figure Description
[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0034] Figure 1 A flowchart for a human-land interaction simulation method that couples a large language model and a process model;
[0035] Figure 2 A structural diagram of a human-land interaction simulation system that couples a large language model and a process model;
[0036] Figure 3 The overall framework for Qwen3 model training and region application;
[0037] Figure 4 The simulation results show the marginal emission reduction / pollution reduction effect brought about by each additional 1 yuan per hectare of subsidy. Detailed Implementation
[0038] like Figure 1 As shown, this invention proposes a method for simulating human-land interaction by coupling a large language model and a process model, comprising the following steps:
[0039] S1. Based on survey and statistical data of farmers in the target area, construct a standardized sample set for training the large language model:
[0040] The input content is formed by combining farmers' characteristic information with policy context information, and the output content is formed by the corresponding changes in fertilization behavior. Training samples in the format of "instruction-input-output" are constructed according to the preset instruction template.
[0041] Interpolation is performed on samples corresponding to adjacent policy scenarios, and perturbations within a preset range can be superimposed on the output results to expand the number of samples and improve the model's generalization ability to unobserved policy scenarios.
[0042] S2. Select the pre-trained large language model, and use the standardized sample set to efficiently fine-tune the parameters of the large language model. By searching for multiple training hyperparameter combinations, determine the target hyperparameter combination for model fine-tuning, and obtain a farmer decision agent model that can generate inference results on changes in fertilization behavior based on the input farmer attributes and policy scenarios.
[0043] The process of determining the combination of multiple training hyperparameters includes: setting different training rounds, batch sizes, gradient accumulation steps, and learning rates, and evaluating the model based on repeatedly partitioned training and test sets, thereby selecting farmer decision-making agent models that meet the preset performance conditions.
[0044] S3. In the reasoning stage, the attributes and preset policy variables of each farmer in the target area are input into the farmer decision-making agent model to obtain the fertilization adjustment results corresponding to each farmer.
[0045] The fertilization adjustment results are allocated to spatial grid cells or plot cells according to preset rules to form a set of management input parameters corresponding to the gridded biogeochemical model.
[0046] S4. The fertilization adjustment results are parameterized and mapped to management input parameters that can be recognized by the gridded biogeochemical model. The parameterization process includes at least: scaling the baseline fertilization amount according to the reduction ratio in the fertilization adjustment results, and establishing a correspondence between the scaled fertilization amount and the spatial grid unit or plot unit.
[0047] S5. Drive the gridded biogeochemical model and run it in parallel within the target area in combination with the management input parameters to simulate the changes in environmental indicators such as NH3 volatilization, non-point source nitrogen (NPSN), N2O and CH4 emissions.
[0048] In practical implementation, according to the above plan, the first step is to obtain farmer survey and statistical data. This data should include at least information such as the farmer's geographical location, family labor force, crop type and planting area, economic factors such as fertilizer input and price, and whether they have received scientific fertilization training. Policy scenario variables (e.g., unit area subsidies, fertilizer price changes) should also be set. Each behavioral record should be organized into a unified "instruction-input-output" sample format, where the input is the farmer's attributes and a single policy question, and the output is the change in fertilization behavior under that policy scenario. A certain perturbation can be added by linear interpolation between adjacent survey points to enhance sample diversity and extrapolation ability.
[0049] Subsequently, the pre-trained Qwen3 large-scale language model was selected as the farmer decision-making agent. A parameter-efficient fine-tuning method (such as LoRA) was used to train the model in a targeted manner, enabling it to generate corresponding fertilization behavior changes given the input of "farmer attributes + policy scenario". Stable models were selected as the inference model through training / test set partitioning and multiple randomized repeated evaluations. In the actual inference stage, the large-scale language model was input with the data of each farmer and policy variables within the target area to obtain the fertilization adjustment results for each farmer. These results were then standardized in units and mapped to management input parameters recognizable by the GDNDC model, establishing a correspondence with spatial grids or plot units, thus forming a dynamic coupling interface of "human management parameters - natural process model input".
[0050] Finally, the GDNDC is driven to run in parallel in the target area, outputting environmental indicators such as ammonia volatilization, agricultural non-point source nitrogen loss, and greenhouse gas emissions. It can also generate statistical and spatial distribution results to jointly predict changes in human fertilization behavior and its environmental effects on the nitrogen cycle under policy scenarios. This coupled framework can express the transmission chain of "policy → behavior → emissions / pollution" in a closed loop within the same system, enabling quantitative assessment and regionally differentiated optimization of agricultural environmental policies.
[0051] like Figure 2 As shown, this invention also protects a human-land interaction simulation system that couples a large language model and a process model, comprising:
[0052] The data preprocessing module is used to acquire survey data and statistical data of farmers in the target area, and to standardize the data to construct a sample set for training a large language model; the data preprocessing module is also used to organize farmer characteristic information and policy context information into a preset input format, and to organize the corresponding fertilization behavior change results into an output format, so as to form standardized data that can be used for model training and inference.
[0053] The farmer decision-making agent module is used to call the pre-trained large language model and perform efficient parameter fine-tuning of the large language model based on the standardized sample set to construct the farmer decision-making agent model; the farmer decision-making agent module is also used to receive farmer attribute information and preset policy variables during the inference stage and output the corresponding fertilization adjustment results of the farmers under the policy scenario.
[0054] The coupling interface module is used to unify and parameterize the fertilization adjustment results and map them into management input parameters that can be recognized by the gridded biogeochemical model. The coupling interface module is also used to scale the baseline fertilization amount according to the fertilization adjustment results and establish a correspondence between the scaled management input parameters and spatial grid units or plot units.
[0055] The gridded biogeochemical model simulation module is used to receive the management input parameters and drive the gridded biogeochemical model to perform grid-by-grid parallel simulations in the target area to obtain environmental indicator changes including NH3 volatilization, non-point source nitrogen (NPSN), N2O emissions and CH4 emissions.
[0056] The results output and analysis module is used to summarize, statistically analyze, and spatially visualize the results of changes in the environmental indicators, and generate the marginal emission reduction and pollution reduction effects and their spatial distribution results under different policy scenarios.
[0057] Example
[0058] like Figure 3 As shown, this embodiment uses Hainan Island as the study area and applies the human-environment interaction simulation framework of this invention. The "Qwen3–GDNDC" coupled simulation and subsidy policy evaluation at a 1 km grid scale in the double-cropping rice area of Hainan Island are presented. Hainan Island, as an important rice-producing area in southern my country, has tropical monsoon climate characteristics: high average annual temperature, abundant rainfall, and a long frost-free period. Maintaining high yields typically requires high-intensity fertilization. However, the warm and humid climate combined with high nutrient input significantly increases the risk of nitrogen loss through ammonia volatilization, denitrification, and runoff, thereby exacerbating agricultural non-point source pollution and affecting regional water quality. Therefore, the double-cropping rice fields of Hainan Island are spatially rasterized at a 1 km resolution, with the simulation period set from 2015 to 2019. Input data includes meteorological, soil, land use, and management measures. Within each grid cell, GDNDC is used to simulate crop growth, soil carbon and nitrogen cycling processes, and nitrogen loss, and outputs indicators such as NH3 volatilization, agricultural non-point source nitrogen loss (NPSN), N2O emissions, and CH4 emissions.
[0059] The coupled operation mode is as follows: First, policy variables (e.g., "subsidy amount per unit area / subsidy increment") and the farmer or management characteristics represented by the grid are used as inputs. Qwen3 generates fertilization adjustment results (e.g., reduction ratio / interval) for the corresponding scenario. Then, the output is unified in units and mapped to management input parameters recognizable by GDNDC (e.g., scaling the baseline fertilization amount according to the reduction ratio). GDNDC is then run cell by cell across the entire island grid, and the results are aggregated to obtain the island-scale changes in pollution / greenhouse gas emissions and their spatial distribution. The results can be interpreted in conjunction with the appendix. Figure 4Description: Using the marginal emission / pollution reduction effect of each additional RMB 1 / ha subsidy as a spatial indicator (k-value), Hainan Island exhibits significant regional heterogeneity—marginal emission reduction from NH3 volatilization is mainly concentrated in the northern and northwestern plains; marginal emission reduction from NPSN loss is more prominent in the central-southern and eastern basins; N2O emission reduction is more significant in the eastern coastal areas, while CH4 emission reduction is stronger in the central and southwestern regions. Overall, the northern and central-southern regions demonstrate more comprehensive emission and pollution reduction benefits across multiple indicators, while the response in the southeastern coastal areas is relatively weaker, suggesting that subsidy policies should be optimized in a zoned and differentiated manner based on local agricultural and environmental conditions.
[0060] In comparison, several alternative technical solutions may exist, but all have shortcomings, making it difficult to achieve the comprehensive effects of this invention. Firstly, using only process models such as DNDC for scenario analysis, where researchers pre-set uniform reduction ratios or fixed management parameters (e.g., "increased subsidies → island-wide reduction of 10% / 20%) and then drive the model to output various indicators, relies on static, homogeneous, and strongly hypothetical human management inputs. This makes it difficult to characterize the different responses of different regions and farmers to the same policy, easily distorting the "policy intensity-behavioral change" relationship and weakening the explanatory power and credibility of policy evaluation. Secondly, constructing a rule base / expert system or agent-based simulation (ABM) involves manually writing decision rules or thresholds based on expert experience and outputting fertilization behavior. However, this approach is highly dependent on experience for rules and parameters, lacks comprehensive coverage, and has high maintenance costs. When policy issues increase or regional differences intensify, rules easily expand and conflict, requiring extensive calibration for stable operation. It also struggles to automatically generate usable decisions from survey texts and complex policy scenarios. In contrast, this invention, through a dynamic coupling link of "generating fertilization adjustments using a large language model—standardizing the mapping to management input—driving GDNDC outputs pollution and greenhouse gas emissions," has advantages in terms of realism, scalability, and the ability to characterize regional differences. Thirdly, traditional econometric or deep learning models (linear regression, Logit / Probit, DID, etc.) are first used to fit the function of "subsidy / price → reduction in fertilization," and then the predicted results are input into the process model. Such methods typically rely on a pre-defined function form, have limited feature interaction capabilities, are sensitive to sample size and representativeness, and often require remodeling or repeated parameter setting for cross-regional migration and multi-policy expansion. Furthermore, the outputs are mostly single-point values or coarse-grained groupings, making it inconvenient to uniformly map them into a scalable sequence of management measures.
[0061] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All modifications made according to the spirit and essence of the main technical solution of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for simulating human-land interaction by coupling a large language model and a process model, characterized in that, Includes the following steps: Based on survey and statistical data of farmers in the target area, a standardized sample set is constructed for training a large language model; A pre-trained large language model is selected, and the parameters of the large language model are fine-tuned using the standardized sample set to obtain a farmer decision-making agent model that can generate inference results on changes in fertilization behavior based on the input farmer attributes and policy scenarios. During the reasoning phase, the attributes and preset policy variables of each farmer in the target area are input into the farmer decision-making agent model to obtain the fertilization adjustment results corresponding to each farmer. The fertilization adjustment results are parameterized and mapped to management input parameters that can be recognized by the gridded biogeochemical model. The parameterization process includes at least: scaling the baseline fertilization amount according to the reduction ratio in the fertilization adjustment results, and establishing a correspondence between the scaled fertilization amount and spatial grid units or plot units. The gridded biogeochemical model is driven to run in parallel within the target area, in conjunction with the management input parameters, to simulate the process of environmental indicator changes.
2. The human-land interaction simulation method coupling a large language model and a process model according to claim 1, characterized in that, Methods for constructing standardized sample sets for training large language models include: The input content is formed by combining farmers' characteristic information with policy context information, and the output content is formed by the corresponding changes in fertilization behavior. Training samples in the format of "instruction-input-output" are constructed according to the preset instruction template.
3. The human-land interaction simulation method coupling a large language model and a process model according to claim 2, characterized in that, Constructing the standardized sample set also includes: Interpolation is performed on samples corresponding to adjacent policy scenarios, and perturbations are added to the output results to expand the sample size and improve the model's generalization ability to unobserved policy scenarios.
4. The human-land interaction simulation method coupling a large language model and a process model according to claim 1, characterized in that, When fine-tuning the parameters of the pre-trained large language model, the target hyperparameter combination for model fine-tuning is determined by searching multiple training hyperparameter combinations.
5. The human-land interaction simulation method coupling a large language model and a process model according to claim 1, characterized in that, After obtaining the fertilizer adjustment results for each farmer, the following is also included: The fertilization adjustment results are allocated to spatial grid cells or plot cells according to preset rules to form a set of management input parameters corresponding to the gridded biogeochemical model.
6. The human-environment interaction simulation method coupled with a large language model and a process model according to claim 1, characterized in that the environmental indicators include NH3 volatilization, non-point source nitrogen (NPSN), N2O and CH4 emissions.
7. A human-land interaction simulation system coupling a large language model and a process model, characterized in that, include: The data preprocessing module is used to acquire survey data and statistical data of farmers in the target area, and to standardize the data to construct a sample set for training a large language model; The data preprocessing module is also used to organize farmer characteristic information and policy scenario information into a preset input format, and organize the corresponding fertilization behavior change results into an output format, so as to form standardized data that can be used for model training and inference. The farmer decision-making agent module is used to call the pre-trained large language model and perform efficient parameter fine-tuning of the large language model based on the standardized sample set to construct the farmer decision-making agent model; the farmer decision-making agent module is also used to receive farmer attribute information and preset policy variables during the inference stage and output the corresponding fertilization adjustment results of the farmers under the policy scenario. The coupling interface module is used to unify and parameterize the fertilization adjustment results and map them into management input parameters that can be recognized by the gridded biogeochemical model. The coupling interface module is also used to scale the baseline fertilization amount according to the fertilization adjustment results and establish a correspondence between the scaled management input parameters and spatial grid units or plot units. The gridded biogeochemical model simulation module is used to receive the management input parameters and drive the gridded biogeochemical model to perform grid-by-grid parallel simulations in the target area to obtain the results of environmental indicator changes. The results output and analysis module is used to summarize, statistically analyze, and spatially visualize the results of changes in the environmental indicators, and generate the marginal emission reduction and pollution reduction effects and their spatial distribution results under different policy scenarios.
8. The human-land interaction simulation system coupling a large language model and a process model according to claim 7, characterized in that, The results of changes in environmental indicators include NH3 volatilization, non-point source nitrogen (NPSN), N2O emissions, and CH4 emissions.