Livestock and poultry manure assimilation capacity evaluation and intelligent matching and scheduling system

Abstract

The present application relates to the field of wisdom agriculture and agricultural resource environment management technology, especially for the livestock and poultry manure disposal capacity evaluation and intelligent matching and scheduling system of the combination of planting and breeding, wherein, the data acquisition and fusion module is used for multi-source heterogeneous data space-time alignment and multi-modal dimension reduction fusion, forming a standardized multi-dimensional feature tensor; the dynamic disposal capacity evaluation module constructs a physical information space-time graph neural network disposal potential field model based on the tensor, embeds the nitrogen and phosphorus leaching and volatile decay partial differential equation residual as a differentiable regular term, and outputs the space-time disposal potential field distribution map through space-time forward deduction; the intelligent matching and scheduling module maps the potential field into a multi-agent scheduling hypergraph, constructs a multi-objective utility function, introduces a potential field-game dual-channel attention mechanism for dynamic weighting, and solves the global optimal scheduling instruction based on the Lagrange hard constraint and differential game algorithm. The present application realizes the precise dynamic evaluation and multi-objective intelligent scheduling closed loop of disposal capacity, and significantly improves the utilization rate of manure resources.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture and agricultural resource and environmental management, specifically to a system for assessing and intelligently matching the disposal capacity of livestock and poultry manure for integrated crop and livestock farming. Background Technology

[0002] Currently, the assessment and scheduling of manure disposal largely rely on manual experience or static nutrient balance tables, which presents the following technical bottlenecks:

[0003] First, the assessment of absorption capacity lacks spatiotemporal dynamic quantitative means. Traditional methods are difficult to couple with the interaction between complex physicochemical processes such as nitrogen and phosphorus leaching and volatilization and regional meteorological and hydrological conditions and soil background, which can easily lead to local environmental overload or idle nutrient resources.

[0004] Second, the matching and scheduling of planting and breeding often adopts single-objective or deterministic optimization models, which do not fully integrate the traffic resistance of logistics network, agronomic application window, carbon emission cost and multi-stakeholder game behavior, resulting in poor scientificity, economy and feasibility of scheduling schemes;

[0005] Third, the systems generally lack real-time feedback on actual field application effects and model self-correction mechanisms, making it easy for prediction results to become disconnected from actual absorption status and difficult to adapt to complex and ever-changing agricultural environments and nonlinear decay patterns.

[0006] Therefore, there is an urgent need for a system that can integrate physical mechanisms and data-driven approaches to achieve dynamic and accurate assessment of absorption capacity, intelligent matching and scheduling of multiple objectives, and closed-loop adaptive correction, in order to solve the problems of mismatch between planting and breeding factors and environmental risk management. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent matching and scheduling system for assessing and matching the capacity of livestock and poultry manure disposal for integrated crop and livestock farming, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An intelligent matching and scheduling system for assessing and matching the disposal capacity of livestock and poultry manure in integrated crop and livestock farming includes:

[0010] The data acquisition and fusion module is used to collect real-time data on the output characteristics of livestock manure, soil-crop background data, regional meteorological and hydrological data, and logistics network data. After spatiotemporal benchmark alignment and multimodal feature dimensionality reduction fusion, a standardized multidimensional feature tensor representing the relationship between planting and breeding elements is formed.

[0011] The dynamic absorption capacity assessment module, connected to the data acquisition and fusion module, is used to construct and run a physical information spatiotemporal graph neural network absorption potential field model based on standardized multidimensional feature tensors. The physical information spatiotemporal graph neural network absorption potential field model constructs a graph topology with spatial discrete grids as nodes and material flux between grids as edges. It captures the spatial diffusion relationship of nutrients through a graph convolutional network and embeds the residuals of the partial differential equations of nitrogen and phosphorus leaching and volatilization attenuation as differentiable regular terms into the network loss function. Through spatiotemporal forward extrapolation, it calculates the nutrient supply and demand potential energy difference of each grid within a preset agricultural time window and outputs a spatiotemporal absorption potential field distribution map containing the dynamic absorption threshold and environmental carrying capacity index.

[0012] The intelligent matching and scheduling module, connected to the dynamic absorption capacity assessment module, maps the spatiotemporal absorption potential field distribution map into a multi-agent scheduling hypergraph. Using farms and planting plots as agent nodes, the edge weights are determined by the joint function of the potential gradient field of the absorption potential field distribution map and the traffic impedance of the logistics network. A multi-objective utility function is constructed, comprising logistics cost, agronomic suitability, environmental overload risk, and carbon emission cost. A potential field-game dual-channel attention mechanism is introduced to extract implicit supply and demand correlation features between nodes. The multi-objective utility function is dynamically weighted to generate a game utility matrix. Based on the Lagrange multiplier method, the environmental carrying capacity index is embedded as a hard constraint into the Nash equilibrium solution process. A differential game adaptive tracking algorithm is used to solve the batch allocation of manure and wastewater transfer and the multi-objective path optimization, outputting the globally optimal scheduling instruction set.

[0013] The closed-loop feedback and adaptive correction module, connected to the intelligent matching and scheduling module and the field in-situ monitoring terminal, is used to receive soil available nutrient inversion data and crop growth indicators after manure application, construct observation residual sequences, compare the observation residual sequences with the predicted states of the physical information spatiotemporal graph neural network absorption potential field model in the corresponding grid and time window, calculate the joint posterior bias distribution through the online Bayesian update mechanism with shared hidden state space, and synchronously and inversely correct the decay dynamic parameters of the partial differential equation and the risk weighting coefficient of the multi-objective utility function, forming a closed-loop iterative control of evaluation-scheduling-feedback-correction.

[0014] As can be seen from the technical solution provided by the present invention above, the beneficial effects of the livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming provided by the present invention are:

[0015] The deep integration of physical mechanisms and data-driven approaches significantly improves the accuracy and physical consistency of absorption assessment: By embedding the residuals of the partial differential equations for nitrogen and phosphorus leaching and volatilization decay as differentiable regular terms into the loss function of the spatiotemporal graph neural network, the shortcomings of pure data-driven models, such as poor generalization ability and easy violation of the law of conservation of matter, are overcome. The model is constrained by physical conservation and dynamic laws in real time during spatiotemporal forward extrapolation, accurately outputting dynamic absorption thresholds and environmental carrying capacity indices, effectively avoiding the risk of local overload.

[0016] A dual-channel attention mechanism based on potential field and game theory, along with differential game equilibrium solution, enables multi-objective global optimal scheduling. This involves innovatively constructing a multi-agent scheduling hypergraph, integrating potential energy gradient compliance and logistics impedance to determine edge weights, and introducing a dual-channel attention mechanism combining potential field physical laws and historical strategy game theory to dynamically weight multi-objective utility functions. Furthermore, the Lagrange multiplier method incorporates environmental carrying capacity as a hard constraint into the generalized Nash equilibrium solution. A differential game adaptive tracking algorithm is employed to simultaneously optimize batch allocation and path optimization, taking into account logistics costs, agronomic suitability, environmental risks, and carbon emission costs.

[0017] Online Bayesian shared latent state correction constructs an evaluation-scheduling-feedback-correction closed-loop iterative control: Based on in-situ field monitoring data, an observation residual sequence is constructed. The deviation of physical attenuation parameters and scheduling risk weight deviation are decoupled through a shared latent state encoder. The joint posterior distribution is updated in real time using online Bayesian recursion with particle filtering. The PDE dynamic parameters and utility function weighting coefficients are corrected inversely at the same time, enabling the system to autonomously adapt to changes in the regional environment and actual field conditions, thus realizing the leap from static planning to dynamic closed-loop intelligent control.

[0018] Multimodal spatiotemporal alignment and heterogeneous Transformer dimensionality reduction fusion solidify the foundation of high-quality data: For multi-source heterogeneous data such as aquaculture, planting, meteorology, and road networks, hierarchical collaborative Kriging interpolation, spatiotemporal local weighted regression, and heterogeneous modal interactive Transformer are used to perform spatiotemporal benchmark alignment and cross-modal feature compression, effectively eliminating dimensional differences and the curse of dimensionality. This provides high-fidelity, low-redundancy standardized multidimensional feature tensor inputs for subsequent graph neural networks and game scheduling models, ensuring the stability and convergence speed of end-to-end decision-making. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of the livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] To better understand the above technical solutions, the following will provide a detailed description of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, this embodiment of the invention provides a system for assessing and intelligently matching the disposal capacity of livestock and poultry manure for integrated crop and livestock farming, including:

[0023] The data acquisition and fusion module is used to collect real-time data on the output characteristics of livestock manure, soil-crop background data, regional meteorological and hydrological data, and logistics network data. After spatiotemporal benchmark alignment and multimodal feature dimensionality reduction fusion, a standardized multidimensional feature tensor representing the relationship between planting and breeding elements is formed.

[0024] The dynamic absorption capacity assessment module, connected to the data acquisition and fusion module, is used to construct and run a physical information spatiotemporal graph neural network absorption potential field model based on standardized multidimensional feature tensors. The physical information spatiotemporal graph neural network absorption potential field model constructs a graph topology with spatial discrete grids as nodes and material flux between grids as edges. It captures the spatial diffusion relationship of nutrients through a graph convolutional network and embeds the residuals of the partial differential equations of nitrogen and phosphorus leaching and volatilization attenuation as differentiable regular terms into the network loss function. Through spatiotemporal forward extrapolation, it calculates the nutrient supply and demand potential energy difference of each grid within a preset agricultural time window and outputs a spatiotemporal absorption potential field distribution map containing the dynamic absorption threshold and environmental carrying capacity index.

[0025] The intelligent matching and scheduling module, connected to the dynamic absorption capacity assessment module, maps the spatiotemporal absorption potential field distribution map into a multi-agent scheduling hypergraph. Using farms and planting plots as agent nodes, the edge weights are determined by the joint function of the potential gradient field of the absorption potential field distribution map and the traffic impedance of the logistics network. A multi-objective utility function is constructed, comprising logistics cost, agronomic suitability, environmental overload risk, and carbon emission cost. A potential field-game dual-channel attention mechanism is introduced to extract implicit supply and demand correlation features between nodes. The multi-objective utility function is dynamically weighted to generate a game utility matrix. Based on the Lagrange multiplier method, the environmental carrying capacity index is embedded as a hard constraint into the Nash equilibrium solution process. A differential game adaptive tracking algorithm is used to solve the batch allocation of manure and wastewater transfer and the multi-objective path optimization, outputting the globally optimal scheduling instruction set.

[0026] The closed-loop feedback and adaptive correction module, connected to the intelligent matching and scheduling module and the field in-situ monitoring terminal, is used to receive soil available nutrient inversion data and crop growth indicators after manure application, construct observation residual sequences, compare the observation residual sequences with the predicted states of the physical information spatiotemporal graph neural network absorption potential field model in the corresponding grid and time window, calculate the joint posterior bias distribution through the online Bayesian update mechanism with shared hidden state space, and synchronously and inversely correct the decay dynamic parameters of the partial differential equation and the risk weighting coefficient of the multi-objective utility function, forming a closed-loop iterative control of evaluation-scheduling-feedback-correction.

[0027] In this embodiment, the data acquisition and fusion module is the sensing center and data base of the livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming. It provides unified and standardized input data for subsequent dynamic disposal capacity assessment and intelligent matching scheduling by real-time acquisition and standardized processing of multi-source heterogeneous data from the entire crop and livestock farming chain.

[0028] The data acquisition and fusion module is primarily responsible for real-time acquisition of data on manure output characteristics from livestock farms, soil-crop baseline data from planting farms, regional meteorological and hydrological data, and logistics network data. It eliminates anomalies and noise in the raw data through data cleaning and standardization, unifies the spatial coordinates and temporal dimensions of all data through spatiotemporal benchmark alignment, and then mines the intrinsic correlations between different data through multimodal feature dimensionality reduction and fusion, ultimately forming a standardized multidimensional feature tensor representing the correlation between planting and livestock elements. This module is the starting point for the system's data flow; the accuracy and standardization of its output data directly determine the precision of subsequent absorption capacity assessments and the rationality of scheduling decisions. The data acquisition and fusion module specifically includes:

[0029] Data cleaning and standardization unit:

[0030] Real-time access to multi-source data: Continuously access data on manure production characteristics from monitoring equipment at the breeding end, soil-crop background data from field monitoring terminals at the planting end, meteorological and hydrological data from regional meteorological and hydrological stations, and logistics network data from geographic information systems to achieve real-time aggregation of core data across the entire planting and breeding chain;

[0031] Anomaly detection and repair: Differentiated processing strategies are adopted for the anomaly characteristics of different types of data; outliers in daily manure removal volume, total nitrogen concentration, and total phosphorus concentration are identified and interpolated for livestock manure output trait data; abnormal fluctuations in nitrogen and phosphorus content and nutrient demand curves during crop growth period are removed from soil-crop baseline data at the planting end; the spatiotemporal consistency of rainfall and temperature series is checked for regional meteorological and hydrological data; and the integrity of road segment passage time and restricted periods is verified for logistics network data, resulting in a cleaned multi-source heterogeneous original dataset.

[0032] Data standardization: The range standardization method is used to unify the dimensions of the cleaned multi-source data. The conversion formula is as follows: ,in, For standardized data values, The original data values, This represents the minimum value of the corresponding data sequence. The maximum value of the corresponding data sequence is used; standardization eliminates the differences in dimensions and numerical magnitudes between different indicators, making various types of data comparable.

[0033] Spatiotemporal reference alignment unit:

[0034] Spatiotemporal aggregation of livestock farm data: The output characteristics data of cleaned livestock farm manure are spatially located according to the center coordinates of the farm, and the discrete sampling data are aggregated into a time series of equal length daily points based on the daily manure production rhythm of the farm, so as to realize the spatiotemporal structured expression of livestock farm manure production data.

[0035] Planting-end data grid mapping: The hierarchical co-kriging interpolation method is used to map the soil-crop background data of discrete sampling points at the planting end to a preset spatial discrete grid. Each grid is assigned the attributes of soil nutrient content and crop nutrient demand during the growth period, so as to realize the spatial continuous expression of the planting end absorption potential data.

[0036] Spatiotemporal unification of multi-source data: Regional meteorological and hydrological data are downscaled to the same temporal resolution of spatial discrete grids through spatiotemporal local weighted regression; logistics network data are converted into time-varying network impedance maps with intersections as nodes and road segment traffic costs as edge weights; all data are unified to the same coordinate reference system and timestamp index to form a spatiotemporally aligned multi-source dataset;

[0037] Multimodal feature dimensionality reduction and fusion unit:

[0038] Heterogeneous modal tokenization: Using spatiotemporally aligned multi-source datasets as input, a heterogeneous modal interactive Transformer network enhanced with location encoding is constructed; the release characteristics of livestock manure, soil absorption potential, meteorological and hydrological driving states and road network accessibility features are mapped to modality-specific tokens, and corresponding spatiotemporal location encodings are injected into each token to preserve the spatiotemporal context information of the data;

[0039] Cross-modal association feature extraction: The attention weights between different modal tokens are calculated through a multi-head cross-modal attention mechanism to capture the supply and demand response relationship between manure supply and soil absorption, the regulatory relationship of meteorological and hydrological factors on nutrient decay, and the constraint pattern of road network impedance on the flow path, thereby generating high-dimensional modal interaction features containing multi-dimensional interaction information.

[0040] Variational feature compression: Under the information bottleneck variational autoencoder framework, high-dimensional modal interaction features are compressed into a low-dimensional latent space. While retaining core feature information, the data dimensionality is greatly reduced, generating a low-dimensional fusion feature vector sequence, avoiding the curse of dimensionality and improving the computational efficiency of subsequent models.

[0041] Tensor construction and output units:

[0042] Spatiotemporal partitioning and stacking: Receive a low-dimensional fused feature vector sequence, and according to the grid number of the spatial discrete grid and the time step of the preset agricultural time window, perform spatiotemporal partitioning and stacking on the low-dimensional fused feature vector sequence to form a three-dimensional feature structure containing spatial dimension, time dimension and feature dimension.

[0043] Layer normalization: The stacked 3D feature structure is subjected to layer normalization channel by channel. The processing formula is as follows: ,in, These are the normalized eigenvalues. For scaling parameters, For offset parameters, For input feature values, The mean of the feature channels. The variance of the characteristic channels. To prevent tiny constants with zero denominators, layer normalization is used to eliminate differences in numerical distribution between different feature channels, thereby improving the training stability of subsequent neural network models.

[0044] Standardized tensor output: Generate standardized multidimensional feature tensors that characterize the relationship between planting and breeding elements, and output them to the dynamic absorption capacity assessment module to provide standardized input data for subsequent absorption capacity assessment;

[0045] Specifically, targeted detection and repair methods were adopted for the abnormal distribution characteristics of different types of data; the 3σ criterion was used to identify outliers in the pollution data from the aquaculture sector, and linear interpolation was used to repair missing and abnormal data; the sliding window method was used to detect abnormal fluctuations in the soil-crop data from the planting sector, and the data was filled in using the mean of nearest neighbor points; the outlier detection method was used for meteorological and hydrological data based on the principle of spatiotemporal neighborhood consistency, and spatiotemporal local weighted regression was used to achieve data interpolation; the missing information in the logistics network data was detected by the rule verification method, and the statistical mean of historical data from the same period was used for supplementation; the range standardization method mapped data of different dimensions to a unified scale through linear transformation, eliminating the impact of dimensional differences on subsequent analysis.

[0046] Spatial projection transformation technology based on geographic information system realizes unified conversion of data in different coordinate systems, and UTC timestamp synchronization mechanism is used to unify the time dimension of all data; hierarchical co-kriging interpolation technology utilizes the spatial autocorrelation of soil nutrients and the co-correlation between different soil layers to realize high-precision mapping of sparse sampling point data to continuous grids; spatiotemporal local weighted regression technology assigns weights to neighboring spatiotemporal points that decay with distance to realize the downscaling of coarse resolution meteorological and hydrological data to fine resolution grids, ensuring complete matching of spatiotemporal resolution of multi-source data.

[0047] A heterogeneous modal interaction network is constructed based on the Transformer architecture. Spatiotemporal location information is injected into tokens of different modalities through positional encoding, enabling the model to perceive the spatiotemporal context of the data. The multi-head cross-modal attention mechanism mines the nonlinear correlation between heterogeneous data from different dimensions by computing the weights of multiple attention heads in parallel, and comprehensively captures the interaction relationship of multiple elements in the farming system. The information bottleneck variational autoencoder introduces information bottleneck constraints to force the model to retain the core features most relevant to subsequent tasks, while filtering redundant information, and achieves effective compression of high-dimensional features.

[0048] The low-dimensional fusion features are structurally stacked according to the two-dimensional index of spatial grid and time window to form a three-dimensional tensor with a clear spatiotemporal structure, so that the subsequent graph neural network model can directly utilize the spatiotemporal correlation information of the data; the layer normalization technique performs mean and variance normalization processing independently on each feature channel, eliminates the numerical distribution offset between different feature channels, accelerates the convergence speed of the subsequent model and improves the generalization ability of the model.

[0049] In this embodiment, the dynamic absorption capacity assessment module is the core assessment engine of the livestock and poultry manure absorption capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming. Based on the standardized multidimensional feature tensor output by the data acquisition and fusion module, it constructs a absorption potential field model that integrates physical prior knowledge and data-driven learning, thereby realizing the spatiotemporal dynamic quantitative assessment of the regional scale livestock and poultry manure absorption capacity.

[0050] The dynamic absorption capacity assessment module is primarily responsible for constructing and running a physical information spatiotemporal graph neural network absorption potential field model. This model uses spatial discrete grids as nodes and inter-grid material flux as edges to construct a graph topology. It captures the spatial diffusion relationship of nutrients through a graph convolutional network and embeds the residuals of the partial differential equations for nitrogen and phosphorus leaching and volatilization decay as differentiable regularization terms into the network loss function. Through spatiotemporal forward extrapolation, it calculates the nutrient supply and demand potential energy difference of each grid within a preset agricultural time window, ultimately outputting a spatiotemporal absorption potential field distribution map containing the dynamic absorption threshold and environmental carrying capacity index. The output of this module is the core basis for the intelligent matching and scheduling module's decision-making regarding the transfer of manure and wastewater. Its assessment accuracy directly determines the scientific validity and environmental safety of the scheduling scheme. The dynamic absorption capacity assessment module specifically includes:

[0051] Directed graph topology building blocks:

[0052] Mesh node extraction: Receive the standardized multidimensional feature tensor output by the data acquisition and fusion module, extract the geographical location identifiers and mesh adjacency relationships corresponding to the spatial discrete mesh, and use each spatial discrete mesh as a graph node to lay the foundation for the subsequent construction of the spatial topology structure of the potential field.

[0053] Directed edge generation: Based on surface hydrological flow direction, topographic gradient and soil texture connectivity, directed edges are constructed to characterize the direction and capacity of nitrogen and phosphorus flux transport between grids; surface hydrological flow direction determines the direction of nutrient transport with surface runoff, topographic gradient affects the driving force of nutrient transport, and soil texture connectivity determines the infiltration and transport capacity of nutrients in soil pores.

[0054] Assigning transport attributes: Assign a transport attribute vector consisting of soil permeability coefficient and nutrient diffusion coefficient to each directed edge; soil permeability coefficient represents the rate of water infiltration in soil, and nutrient diffusion coefficient represents the rate of molecular diffusion of nitrogen and phosphorus nutrients in soil aqueous solution. Together, they determine the efficiency and intensity of nutrient transport between grids.

[0055] Initial node state assignment unit:

[0056] Calculation of manure release potential energy: Extract the manure output trait data from the standardized multidimensional feature tensor, map it after spatiotemporal alignment to obtain the daily release of nitrogen and phosphorus of manure in each grid, and combine it with the manure nutrient mineralization rate curve to calculate the manure nitrogen and phosphorus release potential energy of each grid within the preset agricultural time window, as the node supply state component.

[0057] Soil absorption background potential energy calculation: Analyze the crop background data of the planting end soil in the standardized multidimensional feature tensor, extract the nitrogen and phosphorus storage at different levels of the soil profile of each grid, and combine the soil nutrient availability coefficient to calculate the soil absorption background potential energy as the basic absorption state component of the node.

[0058] Crop absorption potential energy calculation: Based on the nutrient demand curve of crop growth period, calculate the expected absorption of crop nitrogen and phosphorus at different time steps within the preset agricultural time window for each grid. Combine the crop root distribution depth and nutrient absorption efficiency to calculate the crop absorption potential energy as the node demand state component.

[0059] Assignment of hydrothermal driving factors: Temperature, rainfall, soil moisture content and other hydrothermal driving factors obtained from regional meteorological and hydrological data are used as nodal environmental driving state components; hydrothermal factors directly affect microbial activity and nutrient conversion rate and are key environmental variables for regulating the absorption process;

[0060] Initial state vector integration: The nitrogen and phosphorus release potential energy of manure, the background potential energy of soil absorption, the potential energy of crop absorption, and the hydrothermal driving factors are integrated into the initial node state vector of the corresponding graph node, so that the directed graph topology is accompanied by an initial state driven by multiple physical quantities.

[0061] Differentiable physical regularity term building blocks:

[0062] Partial differential equation discretization: The partial differential equations describing the kinetic processes of nitrogen and phosphorus leaching, volatilization and decay are discretized along the spatial dimension of the spatial discretization grid and the time dimension of the preset agricultural time window to construct the physical residual operator in the form of difference; the nitrogen and phosphorus leaching process is described by the convection dispersion equation, the volatilization process is described by the first-order kinetic equation, and the microbial decomposition decay process is described by the Monod equation.

[0063] Construction of differentiable regularization term: Based on the physical residual operator in difference form, a differentiable physical regularization term is constructed; this regularization term can calculate the deviation between the nutrient distribution predicted by the network and the physical dynamic law, and use the deviation as part of the loss function to participate in the backpropagation optimization of the network;

[0064] Loss function embedding: A differentiable physical regularization term is embedded into the loss function of the graph convolutional network to be trained. This ensures that while learning the data distribution characteristics, the network is constrained by physical conservation laws and dynamic principles, preventing the network from generating predictions that violate physical principles. The expression for the joint loss function is: ,in, For the joint loss function value, To fit the loss value to the data, This represents the loss value for the physical regularization term. The weighting coefficient for the physical regularization term;

[0065] Unit for constructing the potential field inference network:

[0066] Edge convolutional layer construction: Edge convolutional layers are built to transmit information about the interaction between edge transmission attributes and node states; the edge convolutional layer can calculate the nutrient transmission flux between grids based on the transmission attribute vector of the edge and the state vector of the adjacent node, thus realizing the spatial propagation of node states;

[0067] Spatial graph attention layer construction: A spatial graph attention layer is built to adaptively aggregate the nutrient absorption features of neighboring nodes; the spatial graph attention layer can assign different attention weights to different neighboring nodes, focusing on neighboring nodes that have a greater impact on the absorption state of the current node, thereby improving the effectiveness of feature aggregation;

[0068] Construction of gated cyclic time coding unit: A gated cyclic time coding unit is introduced to process time-varying meteorological and hydrological driving sequences; the gated cyclic unit can capture long-term dependencies in the time series and realize the time evolution modeling of the potential field.

[0069] Cross-layer skip connection design: Introducing cross-layer skip connections into the network, the local fine features extracted by the shallow network are fused with the global abstract features extracted by the deep network, preserving the multi-scale absorption potential features and improving the model's ability to represent absorption processes at different spatial scales.

[0070] Forward computation structure integration: The edge convolutional layer, spatial graph attention layer and gated recurrent temporal coding unit are stacked in sequence and cross-layer skip connections are added to form a graph neural network forward computation structure with physical regularization guidance;

[0071] Model training and parameter optimization unit:

[0072] Training sample set construction: Collect standardized multidimensional feature tensors and field soil nutrient change sequences generated by the data acquisition and fusion module corresponding to the historical crop-livestock cycle period to construct a supervised training sample set of absorption potential field; each sample contains the input standardized multidimensional feature tensor and the corresponding field soil nutrient change label;

[0073] Joint loss backpropagation: During training, the data fitting loss and physical regularization loss are calculated simultaneously, and the two are weighted and summed to obtain the joint loss function value. The weight parameters of the network are then updated through the backpropagation algorithm. The data fitting loss uses the mean squared error loss function to measure the deviation between the model's predicted value and the measured value.

[0074] Parameter optimization and validation: The Adam optimizer is used to iteratively optimize the network parameters, and the model performance is evaluated using a validation set. Training is stopped when the validation set loss no longer decreases, and the optimal weight parameters of the absorbable potential inference network are obtained, thus completing the construction of the physical information spatiotemporal graph neural network absorbable potential model.

[0075] Spatiotemporal absorption potential field derivation and output unit:

[0076] Initialization of absorption potential field: The standardized multidimensional feature tensor received in real time is segmented and normalized according to the start time of the preset agricultural time window, and loaded into the physical information spatiotemporal graph neural network absorption potential field model that has been trained. Each spatial discrete grid node is assigned the initial value of nitrogen and phosphorus release potential energy of manure, the initial value of soil absorption background potential energy and hydrothermal driving state at the current time, thus completing the spatiotemporal initialization of absorption potential field.

[0077] Multi-step forward extrapolation: Multi-step forward extrapolation is performed along the time axis with a time step within a preset agricultural time window; in each extrapolation step, the edge convolutional layer calculates the nutrient flux between grids based on the transmission attribute vector and the node state at the previous time step, the spatial graph attention layer aggregates the neighbor influence and updates the node state, and at the same time, the residual of the partial differential equation characterized by the differentiable physical regularization term is used to perform real-time constraint correction on the node state increment, so as to obtain the nitrogen and phosphorus mass stock, leaching loss and volatilization decay of each grid in the current step;

[0078] Nutrient supply and demand potential energy difference calculation: After the extrapolation of each time step is completed, the existing soil nutrient content in the node state vector is read and compared with the expected absorption amount of this step obtained by analyzing the crop growth period absorption curve. Combined with the correction coefficient of hydrothermal driving factor on nutrient availability, the difference between nutrient supply potential energy and absorption demand potential energy is calculated grid by grid to form the nutrient supply and demand potential energy difference of this step.

[0079] Calculation of Dynamic Absorption Threshold and Environmental Carrying Capacity Index: The nutrient supply and demand potential energy difference of each grid is compared with a preset safe benchmark potential energy threshold based on soil environmental capacity. When the potential energy difference exceeds the safe benchmark potential energy threshold, the maximum additional nitrogen and phosphorus equivalent of manure that the grid can absorb without exceeding the environmental risk threshold is marked as the dynamic absorption threshold. The ratio of the already occupied absorption amount in the grid to the dynamic absorption threshold is converted into the current environmental carrying capacity index through a nonlinear mapping. The formula for calculating the environmental carrying capacity index is: ,in, For environmental carrying capacity index, This represents the amount of water already consumed within the grid. The dynamic absorption threshold of the grid;

[0080] Spatiotemporal absorption potential field distribution map generation: Collect the nutrient supply and demand potential energy difference, dynamic absorption threshold and environmental carrying capacity index of all grids in all time steps within the preset agricultural time window, perform co-kriging smoothing interpolation along the spatial dimension and temporal aggregation along the temporal dimension to generate a spatiotemporal absorption potential field distribution map with potential energy gradient representing the surplus or deficiency of absorption capacity and threshold boundary identifying the safe absorption space, and output the distribution map to the intelligent matching and scheduling module;

[0081] Among them, the graph neural network technology of physical information fusion combines prior physical knowledge with data-driven deep learning methods, which solves the problem that pure data-driven models have poor generalization ability and are prone to producing prediction results that violate physical common sense in small sample scenarios. By discretizing the partial differential equations describing the migration and transformation process of nitrogen and phosphorus nutrients into differentiable physical regularization terms and embedding them into the network loss function, the model can learn the statistical laws in the data and the dynamic laws of the physical world at the same time during training, which greatly improves the prediction accuracy and physical consistency of the model. The topological structure of graph neural networks is naturally adapted to the adjacency relationship of spatial discrete grids, which can effectively capture the diffusion and transport characteristics of nutrients in space.

[0082] Spatiotemporal absorption potential field modeling abstracts the regional crop-livestock system into a potential field system driven by both nutrient supply potential energy and absorption demand potential energy. The nutrient supply potential energy comes from the nitrogen and phosphorus release of livestock and poultry manure, while the absorption demand potential energy comes from the soil's nutrient carrying capacity and the crop's nutrient absorption requirements. The magnitude of the potential energy difference characterizes the driving force of nutrient flow between grids, while the direction of the potential energy gradient indicates the optimal direction for natural nutrient flow. The dynamic absorption threshold reflects the maximum safe absorption capacity of the grid under specific agricultural seasons and environmental conditions, while the environmental carrying capacity index quantifies the remaining degree of grid absorption capacity, providing a quantitative indicator for the prevention and control of environmental risks associated with manure flow.

[0083] Multi-scale spatiotemporal feature aggregation achieves effective extraction and fusion of multi-scale spatiotemporal features of the absorption process through the synergistic effect of edge convolutional layers, spatial graph attention layers, and gated recurrent temporal coding units. Edge convolutional layers capture nutrient transport features between local grids, while spatial graph attention layers adaptively aggregate neighbor node features from different spatial ranges, realizing multi-scale spatial feature extraction from local to global. Gated recurrent temporal coding units capture the temporal evolution features of the absorption process, enabling multi-scale temporal modeling of short-term fluctuations and long-term trends. Cross-layer skip connections fuse multi-scale features extracted from networks of different depths, further enhancing the model's representational capabilities.

[0084] In this embodiment, the intelligent matching and scheduling module is the decision-making core of the intelligent matching and scheduling system for the assessment of the capacity of livestock and poultry manure disposal in the integrated farming and breeding industry. Based on the spatiotemporal disposal potential field distribution map output by the dynamic disposal capacity assessment module, it constructs an intelligent scheduling model that integrates physical laws and multi-subject game theory to achieve global optimal matching and flow scheduling of livestock and poultry manure between the breeding end and the planting end.

[0085] The intelligent matching and scheduling module is primarily responsible for mapping the spatiotemporal absorption potential field distribution map into a multi-agent scheduling hypergraph. Using livestock farms and planting plots as agent nodes, it determines edge weights using a joint function of the potential energy gradient field of the absorption potential field distribution map and the traffic impedance of the logistics network. It constructs a multi-objective utility function with components including logistics cost, agronomic suitability, environmental overload risk, and carbon emission cost. A dual-channel attention mechanism of potential field and game theory is introduced to extract implicit supply and demand correlation features between nodes. The multi-objective utility function is dynamically weighted to generate a game theory utility matrix. Based on the Lagrange multiplier method, the environmental carrying capacity index is embedded as a hard constraint in the Nash equilibrium solution process. A differential game adaptive tracking algorithm is used to solve the batch allocation of manure and wastewater and the multi-objective path optimization, ultimately outputting the globally optimal scheduling instruction set. This module is a key hub connecting absorption capacity assessment and actual field application; its decision quality directly determines the resource utilization efficiency, economic benefits, and environmental safety of the planting and breeding system. The intelligent matching and scheduling module specifically includes:

[0086] Multi-agent scheduling hypergraph construction unit:

[0087] Intelligent agent node extraction and attribute assignment: Receive the spatiotemporal absorption potential field distribution map output by the dynamic absorption capacity assessment module, extract the grids where the manure release potential energy is higher than the preset pollution generation threshold from the spatial discrete grids covered by the map as intelligent agent nodes for livestock farms, and extract the grids where planting plots with a positive dynamic absorption threshold and in the crop absorption period are located as intelligent agent nodes for planting plots; assign each intelligent agent node a node attribute vector consisting of potential energy value, current nutrient inventory status, geographical coordinates, and functional type label;

[0088] Supply and demand attraction domain division and soft membership allocation: Starting from the intelligent agent node of the breeding farm, streamline tracing is performed along the potential energy gradient vector field in the spatiotemporal absorption potential field distribution map. The intelligent agent node of the planting plot that the gradient streamline crosses and whose absorption demand potential energy exceeds the demand response threshold is included in the supply and demand attraction domain of the breeding farm. For the intelligent agent node of the planting plot where multiple attraction domains overlap, soft membership is allocated according to the potential energy difference ratio of each upstream breeding farm node to form a many-to-many supply and demand candidate relationship set.

[0089] Scheduling hyperedge definition and attribute assignment: Scheduling hyperedges are defined using a single batch scheduling of manure and sewage transfer as a unit. Each scheduling hyperedge connects a source node of a livestock farm to at least one target node of a planting plot in a multi-to-multi supply and demand candidate relationship set. The total nitrogen and phosphorus equivalent load of the batch, the cumulative attenuation of the potential energy gradient along the path, and the time-varying shortest travel time extracted based on logistics network data are used as hyperedge attributes. All nodes, supply and demand candidate relationship edges, and scheduling hyperedges are combined into a multi-agent scheduling hypergraph.

[0090] Multi-objective utility function building blocks:

[0091] Path joint weight calculation: For each pair of candidate connections between farm agent nodes and planting plot agent nodes in the multi-agent scheduling hypergraph, the potential energy gradient vector sequence of the canonical potential field distribution map on the corresponding transportation path of the candidate connection is extracted. At the same time, the time-varying traffic impedance sequence of the same path is extracted from the logistics network data. The projection modulus of the potential energy gradient vector on the tangential direction of the path is used as the nutrient transport compliance component, and the equivalent value of the traffic impedance after time cost is used as the transport obstacle component. The entropy weight method is used to fuse the compliance component and the obstacle component to obtain the path joint weight. The environmental carrying capacity index attenuation penalty term of the passing grid is embedded in the path joint weight to form the edge weight that characterizes the comprehensive cost of manure transfer.

[0092] Multi-objective component definition and standardization: Logistics cost is defined as the product of edge weight and unit distance transportation cost rate; agronomic suitability is defined as the sliding convolution matching degree of the manure nutrient release time curve and the crop growth period nutrient absorption time curve of the planting plot; environmental overload risk is defined as the conditional risk probability that the nitrogen and phosphorus stock of the planting plot exceeds the dynamic absorption threshold after receiving manure distribution; carbon emission cost is defined as the product of the carbon emission factor of the transportation vehicle and the driving mileage with load; range standardization is performed on logistics cost, agronomic suitability, environmental overload risk and carbon emission cost respectively.

[0093] Multi-objective utility function aggregation: The four types of standardized loss values ​​are synthesized into a multi-objective utility function with the allocation amount of feces and sewage for each candidate connection as the independent variable through a weighted Chebyshev aggregation function. The function expression is as follows: ,in, For breeding farm nodes With planting plot node The amount allocated between them is The overall utility value at that time For the first The initial weights of each target component, For the first The original values ​​of each target component. For the first The minimum value of each target component. For the first The maximum value of each target component;

[0094] Potential Field-Game Theory Dual-Channel Attention Mechanism Unit:

[0095] Potential field attention weight matrix generation: The node attribute vectors and hyperedge attributes in the multi-agent scheduling hypergraph are encoded into query vectors, key vectors and value vectors, and the relative position encoding derived from the potential energy gradient field is injected; the nutrient supply and demand potential energy complementarity intensity between the aquaculture farm node and the planting plot node is calculated by scaling dot product attention, and the potential field attention weight matrix is ​​generated to highlight the supply and demand pairing with large potential energy drop and low transmission loss.

[0096] Game attention weight matrix generation: Taking the agents of each farm and the agents of the planting plot as game participants, extract the historical strategy payoff time difference sequence of each participant in the previous scheduling rolling window; use the causal mask multi-head self-attention mechanism to learn the implicit competition-cooperation time dependence among participants in the non-cooperative elimination game, and generate the game attention weight matrix to express the priority elimination tendency of each participant towards other participants.

[0097] Dynamic weighted coefficient matrix fusion: The potential field attention weight matrix and the game attention weight matrix are fused element-wise and normalized to obtain a dynamic weighted coefficient matrix; the dynamic weighted coefficient matrix is ​​used to adaptively scale the utility components of each farm-plot pair in the multi-objective utility function to generate a game utility matrix; the elements of this game utility matrix represent the comprehensive utility of unit manure equivalent transfer under the condition of simultaneously reflecting the physical absorption law and the multi-agent strategy preference;

[0098] Nash equilibrium solution and scheduling instruction generation unit:

[0099] Hard constraint embedding and generalized Nash equilibrium transformation: The environmental carrying capacity index of all planting plots at each time is not lower than the preset environmental safety lower limit as a rigid constraint, and the Lagrange multiplier vector corresponding to the plot constraint is introduced; the constrained multi-agent scheduling optimization problem is transformed into a generalized Nash equilibrium problem in the form of augmented Lagrange function, and the constraint violation is incorporated into the local utility function of each agent in the form of a penalty term.

[0100] Differential game adaptive pursuit solution: Based on the first-order optimality condition of the generalized Nash equilibrium problem, a system of differential algebraic equations is constructed. The differential game adaptive pursuit algorithm is adopted, and the equilibrium trajectory is pursued by numerical integration along the direction of the synthesis of utility gradient and constraint gradient with the current scheduling state as the initial value. At each time step, the batch allocation of manure, the flow of transportation path and the Lagrange multiplier are updated synchronously until the allocation state converges to the global Nash equilibrium point.

[0101] Global optimal scheduling instruction set compilation: The converged batch loading capacity of the farm, target plot, transportation route node sequence, departure time window and segmented allocation are compiled into a global optimal scheduling instruction set and output; The scheduling instruction set also includes environmental risk warning information and agronomic application suggestions for each batch to guide actual field operations;

[0102] Among them, the multi-agent scheduling hypergraph modeling technology breaks through the limitation of traditional graph models that can only express binary relationships. By introducing scheduling hyperedges, it realizes the direct modeling of the many-to-many relationship in a batch of manure transfer, where a farm simultaneously distributes manure to multiple planting plots. The hyperedge attributes can fully characterize the overall features of batch scheduling, including total load, total transportation time, and total environmental impact, enabling the model to perform global optimization at the batch level. This avoids the optimization fragmentation problem caused by traditional binary graph models that split batches into multiple independent connections. The supply and demand attraction domain partitioning and soft membership allocation mechanism are based on the physical laws of the absorption potential field, effectively reducing the search space of scheduling optimization and improving the solution efficiency.

[0103] The dual-channel attention mechanism of potential field and game theory simultaneously integrates the objective laws of the physical world with the subjective preferences of multi-agent decision-making, achieving a comprehensive extraction of supply and demand relationship characteristics. The potential field attention channel, based on the potential energy gradient and logistics impedance of the absorption potential field, mines the explicit supply and demand relationships between nodes based on physical transmission laws, ensuring that the scheduling scheme conforms to the optimal direction of natural nutrient flow. The game theory attention channel, based on historical strategy payoff data, learns the implicit competition and cooperation relationships of multiple agents in non-cooperative games, enabling the scheduling scheme to better adapt to the decision preferences of each participating agent and improve the feasibility of the scheme. The attention weights of the two channels are fused element by element, achieving an organic unity of physical rationality and agent acceptance.

[0104] The differential game Nash equilibrium solution technique with hard constraints embeds the rigid constraint of environmental carrying capacity into the game solution process through the Lagrange multiplier method, ensuring that no scheduling scheme will exceed the environmental safety threshold. The differential game adaptive tracking algorithm transforms the discrete scheduling decision problem into a continuous differential algebraic equation system solution problem, which can efficiently handle the dynamic scheduling problem of large-scale multi-agent systems. By performing numerical integration along the synthesis direction of utility gradient and constraint gradient, this algorithm can quickly track the equilibrium trajectory of the system, realize the real-time dynamic update of the scheduling scheme, and adapt to the dynamic changes of the absorption potential field over time.

[0105] In this embodiment, the closed-loop feedback and adaptive correction module is the closed-loop control hub and autonomous evolution engine of the livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming. It receives feedback data on the actual application effect in the field, quantifies the deviation between the prediction of the assessment model and the actual operation, and corrects the core parameters of the dynamic disposal capacity assessment model and the intelligent matching scheduling model in reverse, forming a complete closed-loop iterative control loop for assessment, scheduling, feedback, and correction.

[0106] The closed-loop feedback and adaptive correction module is mainly responsible for receiving soil available nutrient inversion data and crop growth indicators after manure application, and constructing an observation residual sequence. It compares the observation residual sequence with the predicted state of the physical information spatiotemporal graph neural network absorption potential field model at the corresponding grid and time window, calculates the joint posterior bias distribution through an online Bayesian update mechanism with a shared hidden state space, and synchronously and inversely corrects the decay dynamic parameters of the partial differential equation and the risk weighting coefficient of the multi-objective utility function. This module is the core guarantee for the system to achieve autonomous learning and continuous optimization, effectively solving the problems of poor regional adaptability of model parameters and the disconnect between scheduling decisions and actual results, ensuring the long-term stable operation of the system in complex and ever-changing agricultural environments. The closed-loop feedback and adaptive correction module specifically includes:

[0107] Gridded field observation sequence generation unit:

[0108] Multi-source field observation data access: Receives real-time data from field in-situ monitoring terminals such as soil ion-selective electrodes, near-infrared spectral probes, and multispectral canopy imagers deployed in planting plots, including soil available nitrogen content, soil available phosphorus content, soil moisture content, soil temperature, crop leaf area index, and normalized vegetation index.

[0109] Bilinear spatiotemporal matching: Bilinear spatiotemporal matching is performed on multi-source observation data according to timestamps and spatial discrete grids, mapping the observation data of discrete sampling points to a unified spatial discrete grid and time step, generating a gridded field observation snapshot sequence;

[0110] Anomaly detection and smoothing: Anomaly jump detection based on adaptive thresholds is performed on the gridded field observation snapshot sequence to identify and remove abnormal data points caused by sensor failure or environmental interference; missing data is filled in using a local weighted scatter smoothing method to ensure the continuity of the observation sequence.

[0111] Kalman Filter State Estimation: The time-domain Kalman filter is used to recursively estimate the soil available nutrient sequence and crop growth index sequence, filter out observation noise, and output the smoothed soil available nutrient state estimation sequence and crop growth status state estimation sequence.

[0112] Observation residual sequence construction unit:

[0113] Crop nutrient uptake inversion: The pre-set crop nutrient uptake inversion model is called to convert the smoothed crop growth status estimation sequence into the actual nitrogen and phosphorus uptake flux of crops at each time step. The crop nutrient uptake inversion model is constructed based on the correlation between crop growth model and multispectral remote sensing data, and can accurately invert the nutrient uptake rate of crops at different growth stages.

[0114] Calculation of measured net soil nutrient storage variables: Based on the smoothed soil available nutrient status estimation sequence and the actual nitrogen and phosphorus uptake flux of crops, the measured soil nutrient storage variable sequence is obtained through grid mass conservation calculation; the mass conservation calculation takes into account the input, output, leaching and volatilization losses of soil nutrients to ensure the accuracy of the calculation results;

[0115] Construction of the interference-free reference absorption trajectory: Based on the baseline value of soil nutrients in the grid before the execution of the scheduling command and the theoretical nutrient input contained in the batch allocation plan of manure and sewage output by the intelligent matching scheduling module, and combined with the predicted attenuation law of the physical information spatiotemporal graph neural network absorption potential field model, an interference-free reference absorption trajectory is constructed; this trajectory represents the expected change process of soil nutrients under ideal conditions.

[0116] Observational residual sequence generation: The deviation between the measured net soil nutrient storage variable sequence and the undisturbed reference absorption trajectory is calculated hourly to generate an observational residual sequence that represents the deviation between the actual absorption effect and the planned absorption target; the observational residual sequence is the core basis for subsequent model deviation analysis and parameter correction;

[0117] Shared hidden state deviation extraction unit:

[0118] Model prediction state extraction: The predicted soil nutrient stock sequence, predicted nitrogen and phosphorus leaching flux sequence, and predicted volatilization decay flux sequence generated by the spatiotemporal graph neural network absorption potential field model in the same spatial discrete grid and the same feedback time window corresponding to the observed residual sequence are used as the model prediction state.

[0119] Construction of normalized innovation sequence: The observation residual sequence is aligned with the model prediction state in time step, and the model prediction state is corrected and restored using the observation residual sequence to construct a normalized innovation sequence that reflects the model prediction bias; the innovation sequence eliminates the influence of observation noise and inherent system fluctuations, and only retains the prediction bias of the model itself;

[0120] Construction of a shared hidden state encoder: A shared hidden state encoder consisting of long short-term memory units and a graph attention layer is constructed. This encoder can extract features of physical model parameter deviation and scheduling model weight deviation at the same time, while preserving the original hidden state space topology of the physical information spatiotemporal graph neural network absorptive potential field model.

[0121] Dual-bias hidden state extraction: The normalized innovation sequence is input into the shared hidden state encoder, which outputs a first hidden state vector representing the deviation of the decay dynamic parameters of the partial differential equation, and a second hidden state vector representing the deviation of the weighted coefficient of environmental overload risk in the multi-objective utility function. The two hidden state vectors decouple the physical deviation from the scheduling deviation, laying the foundation for subsequent independent correction.

[0122] Joint posterior parameter correction unit:

[0123] Particle filter initialization: Using the first hidden state vector and the second hidden state vector as observation variables, initialize a set of weighted random particles in the decay dynamics parameter space and the risk weighting coefficient space respectively; each particle represents a set of possible parameter values, and the weight of the particle represents the probability of that set of parameter values;

[0124] Sequential importance resampling: Employing an online Bayesian recursive framework using particle filtering, the joint posterior probability distribution of decay kinetic parameters and the joint posterior probability distribution of risk-weighted coefficients are recursively calculated through importance resampling and sequential updates. The particle weight update formula is as follows: ,in, For the first Time of the first The weight of each particle, For the first Time of the first The weight of each particle, To observe the likelihood function, For the first Time of the first The state of each particle For the first The observed value at time;

[0125] Maximum a posteriori (MAP) estimation calculation: The expected value of the joint posterior probability distribution of the decay kinetic parameters and the expected value of the joint posterior probability distribution of the risk weighting coefficients are used as the MAP estimate of the joint posterior bias distribution; the MAP estimate represents the most likely parameter bias value under the current observation data.

[0126] Simultaneous correction of dual-model parameters: Based on the expected distribution of the joint posterior bias of the decay kinetic parameters, the microbial decomposition rate constant, ammonia volatilization coefficient, and denitrification rate coefficient in the partial differential equations of nitrogen and phosphorus leaching and volatilization decay are adjusted inversely, and the adjusted parameters are written into the calculation graph of the differentiable physical regularization term of the physical information spatiotemporal graph neural network absorption potential field model, and the model parameters of the dynamic absorption capacity assessment module are updated synchronously; at the same time, based on the expected distribution of the joint posterior bias of the risk weighting coefficients, the weighting factor corresponding to the environmental overload risk component of the multi-objective utility function in the intelligent matching and scheduling module is corrected.

[0127] Closed-loop iterative triggering: After the parameter correction is completed, the absorption potential field is re-evaluated using the updated physical information spatiotemporal graph neural network absorption potential field model, and the next round of intelligent matching scheduling is triggered, forming a closed-loop iterative control loop from evaluation, scheduling, field execution, feedback monitoring to model correction.

[0128] Among them, the multi-source field observation fusion technology combines the advantages of high-precision point observation by in-situ sensors and area observation by multispectral remote sensing. It realizes the mapping of discrete observation data to continuous grid through bilinear spatiotemporal matching. The temporal Kalman filter effectively filters out observation noise by recursively estimating the uncertainty of the fused multi-source observation data, thereby improving the accuracy and stability of the observation data. The crop nutrient absorption inversion technology is based on the quantitative relationship between canopy spectral characteristics and nutrient absorption rate during crop growth. It can non-destructively and in real-time invert the actual nutrient absorption of crops, providing a direct basis for evaluating the absorption effect.

[0129] The principle of mapping observation residuals to model biases constructs an interference-free reference absorption trajectory, comparing the actual observed changes in soil nutrients with the expected changes under ideal conditions. This transforms complex system operational biases into a quantifiable sequence of observation residuals. The interference-free reference absorption trajectory comprehensively considers the initial background conditions, theoretical inputs, and the decay law of model predictions, accurately characterizing the system's operational state in the absence of model biases and external disturbances. The observation residual sequence directly reflects the degree of deviation between model predictions and actual operation, establishing a quantitative mapping relationship between observation data and model parameter biases.

[0130] The bias decoupling principle of the shared hidden state space utilizes a shared hidden state encoder constructed from long short-term memory units and graph attention layers. While preserving the original model's hidden state space topology, it decouples the attenuation dynamics parameter bias of the physical model and the risk weighting coefficient bias of the scheduling model into two independent hidden state vectors. The shared hidden state space ensures that parameter correction does not destroy the model's previously learned knowledge and physical laws, while bias decoupling enables independent correction of physical parameters and scheduling weights, avoiding mutual interference between the two types of parameter corrections and improving the accuracy and stability of the correction.

[0131] The online Bayesian parameter adaptive update principle is based on the particle filter algorithm, which can handle nonlinear and non-Gaussian parameter distribution problems and is suitable for parameter adaptive correction of complex agricultural systems. Particle filtering approximates the posterior probability distribution of parameters by maintaining a set of weighted random particles, and continuously updates the weights and positions of particles through sequential importance resampling to track the dynamic changes of parameters in real time. This method does not require a large amount of historical training data and can update parameters online based on real-time observation data, enabling the model to quickly adapt to the dynamic changes of regional planting and breeding conditions and achieve autonomous evolution of the system.

[0132] In this embodiment, the livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming adopts a closed-loop iterative control architecture. It operates cyclically according to a complete process of initialization, data perception, capacity assessment, intelligent decision-making, execution feedback, and adaptive correction, achieving dynamic and accurate matching and sustainable disposal of livestock and poultry manure resources. The following are the detailed operation steps of the entire system process:

[0133] I. System Global Initialization Phase:

[0134] After the system is powered on and started, it sequentially completes hardware self-tests and software loading for the data acquisition and fusion module, dynamic absorption capacity assessment module, intelligent matching and scheduling module, and closed-loop feedback and adaptive correction module to confirm that each module is operating normally.

[0135] Establish communication connections with monitoring equipment at the aquaculture end, in-situ monitoring terminals at the planting end, regional meteorological and hydrological stations, geographic information systems, and logistics scheduling terminals, and test the stability and real-time performance of the data transmission links;

[0136] Load the system's global configuration parameters, including spatial discrete grid resolution, preset agricultural time window range, environmental safety lower limit threshold, initial parameters of physical dynamic equations, initial weights for multi-objective optimization, and pre-trained weights for each neural network model;

[0137] Initialize the system database and cache space, load historical crop-livestock cycle data, soil environmental baseline data, crop growth period standard curves and logistics network basic data, and complete all preparations before system operation;

[0138] II. Multi-source data acquisition and fusion stage:

[0139] The data acquisition and fusion module collects four types of core data in real time and in parallel: daily manure removal volume, total nitrogen concentration, total phosphorus concentration, and other manure production characteristics data from the livestock end; soil-crop background data such as soil profile stratified nitrogen and phosphorus content and crop growth stage from the planting end; meteorological and hydrological data such as regional rainfall, temperature, wind speed, and surface runoff; and logistics network data such as road passage time, restricted traffic periods, and road grade.

[0140] The collected multi-source heterogeneous data were classified, cleaned, and standardized; outliers in the aquaculture data were identified using the 3σ criterion and repaired using linear interpolation; abnormal fluctuations in the planting data were removed using the sliding window method and filled using the mean of neighboring points; outliers in meteorological and hydrological data were corrected through spatiotemporal consistency checks; missing information in the logistics network was supplemented through rule verification; and the dimensions of all data were unified using the range standardization method.

[0141] Perform spatiotemporal reference alignment; aggregate aquaculture data into daily time series of spatial points based on farm coordinates and daily production and pollution rhythms; map discrete sampling data from the planting end to a spatial discrete grid using hierarchical co-kriging interpolation; downscale meteorological and hydrological data to a uniform grid resolution using spatiotemporal local weighted regression; convert logistics network data into time-varying network impedance maps with intersections as nodes; unify all data to the WGS84 coordinate reference system and UTC timestamp index;

[0142] A heterogeneous modal interactive Transformer network enhanced with location encoding is constructed, which maps the release characteristics of livestock manure, soil absorption potential, meteorological and hydrological driving states, and road network accessibility features into modality-specific tokens; and a multi-head cross-modal attention mechanism is used to capture the nonlinear correlation between multiple elements and generate high-dimensional modal interaction features.

[0143] Under the information bottleneck variational autoencoder framework, high-dimensional features are compressed into a low-dimensional latent space to generate a low-dimensional fusion feature vector sequence. The feature vectors are spatiotemporally partitioned and stacked according to the spatial grid number and the preset agricultural time window. After normalization processing by feature channel layer, a standardized multi-dimensional feature tensor is formed and output to the dynamic absorption capacity evaluation module.

[0144] III. Spatiotemporal Assessment Phase of Dynamic Absorption Capacity:

[0145] The dynamic absorption capacity assessment module receives a standardized multidimensional feature tensor, extracts the geographic location identifiers and adjacency relationships of the spatial discrete grid, and uses each grid as a graph node. Based on the surface hydrological flow direction, topographic gradient and soil texture connectivity, it constructs directed edges that characterize the material flux between grids, and assigns a transport attribute vector composed of soil permeability coefficient and nutrient diffusion coefficient to each edge to form a directed graph topology.

[0146] Calculate the initial state vector of each grid node; extract the daily release of nitrogen and phosphorus from manure and combine it with the mineralization rate curve to obtain the release potential energy of manure; extract the current storage of nitrogen and phosphorus in the soil and combine it with the availability coefficient to obtain the soil absorption background potential energy; obtain the crop absorption potential energy based on the crop growth period demand curve and absorption efficiency; use temperature, rainfall, and soil moisture content as hydrothermal driving factors; integrate the above components into the initial node state vector.

[0147] The partial differential equations describing the nitrogen and phosphorus leaching, volatilization, and decay processes are discretized along the spatiotemporal dimension to construct a difference form physical residual operator; differentiable physical regularization terms are constructed with the physical residual operator as the core, and embedded into the loss function of the graph convolutional network to form a joint loss function;

[0148] Load the pre-trained physical information spatiotemporal graph neural network absorption potential field model, input the directed graph topology and initial node state vector into the model; perform multi-step forward deduction along the time axis with a preset time step; in each deduction step, the edge convolutional layer calculates the nutrient flux between grids, the spatial graph attention layer aggregates neighbor features to update the node state, and the differentiable physical regularization term performs physical constraint correction on the node state increment to obtain the nitrogen and phosphorus stock, leaching loss and volatilization decay of each grid in the current step;

[0149] The difference between the nutrient supply potential energy and the absorption demand potential energy at the current time step is calculated grid by grid to form the nutrient supply and demand potential energy difference; the potential energy difference is compared with the soil environmental capacity safety benchmark threshold to calculate the dynamic absorption threshold and environmental carrying capacity index of each grid.

[0150] Collect the calculation results of all grids at all time steps within the preset agricultural time window, perform co-kriging smoothing interpolation along the spatial dimension, perform temporal aggregation along the temporal dimension, generate a spatiotemporal absorption potential field distribution map, and output it to the intelligent matching and scheduling module.

[0151] IV. Multi-objective intelligent matching and scheduling stage:

[0152] The intelligent matching and scheduling module receives the spatiotemporal absorption potential field distribution map, extracts the intelligent agent nodes of the farm where the release potential energy of manure is higher than the pollution generation start threshold, and the intelligent agent nodes of the planting plot with a positive dynamic absorption threshold and in the crop absorption period, and assigns a corresponding attribute vector to each node.

[0153] Starting from the farm node, streamline tracing is performed along the potential energy gradient vector field to divide the supply and demand attraction domains of each farm. For planting plot nodes where the attraction domains overlap, soft membership degrees are assigned according to the proportion of potential energy difference of the upstream farm to form a many-to-many supply and demand candidate relationship set.

[0154] A scheduling hyperedge is defined with one batch of manure and sewage transfer as the unit. Each hyperedge connects a farm node and at least one planting plot node. The hyperedge is assigned the total nitrogen and total phosphorus equivalent load, potential energy gradient cumulative decay amount and time-varying shortest travel time attributes to construct a multi-agent scheduling hypergraph.

[0155] Calculate the joint path weight for each pair of candidate connections; extract the potential gradient projection modulus on the path as the nutrient transport compliance component, extract the time-varying traffic impedance conversion value as the transport obstruction component, and use the entropy weight method to fuse the two types of components and embed the environmental carrying capacity index decay penalty term to obtain the edge weight;

[0156] The four objective components of logistics cost, agronomic suitability, environmental overload risk and carbon emission cost are calculated separately. After range standardization, a multi-objective utility function is constructed by weighted Chebyshev aggregation function.

[0157] A dual-channel attention mechanism of potential field and game theory is introduced; the attributes of nodes and hyperedges are encoded as query vectors, key vectors and value vectors, and relative position encoding is injected to calculate the potential field attention weight matrix; the historical strategy payoff sequence is input, and the game attention weight matrix is ​​calculated through causal mask multi-head self-attention; the two matrices are fused and normalized element-wise to obtain a dynamic weighted coefficient matrix, and the multi-objective utility function is adaptively scaled to generate the game utility matrix;

[0158] With the environmental carrying capacity index of all planting plots not lower than the environmental safety lower limit as a rigid constraint, the Lagrange multiplier vector is introduced to transform the constrained optimization problem into a generalized Nash equilibrium problem in the form of an augmented Lagrange function.

[0159] Based on the first-order optimality condition, a system of differential algebraic equations is constructed. A differential game adaptive pursuit algorithm is adopted, and numerical integration is performed along the direction of the synthesis of utility gradient and constraint gradient with the current scheduling state as the initial value. The batch allocation of manure, the flow of transportation path and the Lagrange multiplier are updated synchronously until the global Nash equilibrium point is converged.

[0160] The convergence results are compiled into a globally optimal scheduling instruction set that includes the batch loading capacity of the farm, the target plot, the sequence of transportation route nodes, the departure time window, and the segmented allocation. This set is then sent to the logistics scheduling terminal and all participating entities. At the same time, the scheduling plan is synchronized to the closed-loop feedback and adaptive correction module.

[0161] V. Field Implementation and Feedback Monitoring Phase:

[0162] Each participating entity executes manure transfer and field application operations according to the dispatch instruction set, and the logistics dispatch terminal provides real-time feedback on the operation progress and location information;

[0163] Soil ion-selective electrodes, near-infrared spectral probes, and multispectral canopy imagers deployed in the planting plots are used to collect real-time field observation data such as soil available nitrogen content, soil available phosphorus content, soil moisture content, soil temperature, crop leaf area index, and normalized vegetation index.

[0164] The closed-loop feedback and adaptive correction module receives field observation data and performs bilinear spatiotemporal matching based on timestamps and spatial discrete grids to generate a gridded field observation snapshot sequence.

[0165] Anomaly jump detection based on adaptive threshold is performed on the observed sequence to remove abnormal data points; missing data is filled using a local weighted scatter smoothing method; and the soil available nutrient sequence and crop growth sequence are recursively estimated using time-domain Kalman filtering to output a smoothed state estimation sequence.

[0166] VI. Adaptive Correction and Closed-Loop Iteration Stage:

[0167] The pre-set crop nutrient uptake inversion model is invoked to convert the crop growth status estimation sequence into the actual nitrogen and phosphorus uptake flux of crops at each time step; combined with the soil available nutrient status estimation sequence, the measured soil nutrient net storage variable sequence is obtained through grid mass conservation accounting.

[0168] Based on the baseline values ​​of soil nutrients and the theoretical input of manure before scheduling execution, and combined with the decay law predicted by the physical model, an interference-free reference absorption trajectory is constructed; the deviation between the measured net soil nutrient storage variable sequence and the reference trajectory is calculated step by step to generate the observation residual sequence.

[0169] The predicted soil nutrient stock, leaching flux and volatilization decay flux sequences of the spatiotemporal graph neural network absorption potential field model within the corresponding grid and time window are extracted as the model prediction state; the observation residual sequence is aligned with the time step of the model prediction state to construct a normalized information sequence.

[0170] The normalized information sequence is input into a shared hidden state encoder consisting of long short-term memory units and graph attention layers to extract the first hidden state vector representing the deviation of the decay dynamic parameters of the partial differential equation, and the second hidden state vector representing the deviation of the weighted coefficient of the environmental overload risk of the multi-objective utility function.

[0171] Using two hidden state vectors as observed variables, online Bayesian recursion of particle filtering is initiated; a weighted random particle set is maintained in the decay dynamics parameter space and the risk weighting coefficient space, and the joint posterior probability distribution of the two types of parameters is recursively calculated through importance resampling and sequential updates.

[0172] The expected value of the joint posterior probability distribution of the two types of parameters is taken as the maximum a posteriori estimate. The microbial decomposition rate constant, ammonia volatilization coefficient and denitrification rate coefficient in the partial differential equation of nitrogen and phosphorus leaching and volatilization are adjusted in reverse. The parameters of the physical information spatiotemporal graph neural network absorption potential field model are updated simultaneously. At the same time, the weighting factor of the environmental overload risk component in the multi-objective utility function is corrected.

[0173] After the parameter correction is completed, the dynamic absorption capacity assessment module is triggered to recalculate the latest spatiotemporal absorption potential field distribution map, and the system automatically enters the next round of data acquisition, assessment, scheduling, feedback and correction cyclic iteration process;

[0174] VII. System shutdown phase:

[0175] Upon receiving a system stop command or completing all scheduling tasks for a preset planting and breeding cycle, the system sequentially stops data acquisition, model deduction, and scheduling optimization operations of each module.

[0176] Close all communication connections with external devices and terminals, and write all observation data, scheduling records, model parameters and calibration logs for this operation cycle into the system database for persistent storage;

[0177] Generate a system operation summary report, and compile statistics on key indicators such as total amount of sewage disposal, resource utilization rate, environmental risk incidence rate, and logistics costs to provide data support for subsequent system optimization and management decisions;

[0178] Release system memory and computing resources to complete the system shutdown process.

[0179] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for assessing and intelligently matching the disposal capacity of livestock and poultry manure for integrated crop and livestock farming, characterized by: include: The data acquisition and fusion module is used to collect real-time data on the output characteristics of livestock manure, soil-crop background data, regional meteorological and hydrological data, and logistics network data. After spatiotemporal benchmark alignment and multimodal feature dimensionality reduction fusion, a standardized multidimensional feature tensor representing the relationship between planting and breeding elements is formed. The dynamic absorption capacity assessment module, connected to the data acquisition and fusion module, is used to construct and run a physical information spatiotemporal graph neural network absorption potential field model based on standardized multidimensional feature tensors. The physical information spatiotemporal graph neural network absorption potential field model constructs a graph topology with spatial discrete grids as nodes and material flux between grids as edges. It captures the spatial diffusion relationship of nutrients through a graph convolutional network and embeds the residuals of the partial differential equations of nitrogen and phosphorus leaching and volatilization attenuation as differentiable regular terms into the network loss function. Through spatiotemporal forward extrapolation, it calculates the nutrient supply and demand potential energy difference of each grid within a preset agricultural time window and outputs a spatiotemporal absorption potential field distribution map containing the dynamic absorption threshold and environmental carrying capacity index. The intelligent matching and scheduling module, connected to the dynamic absorption capacity assessment module, maps the spatiotemporal absorption potential field distribution map into a multi-agent scheduling hypergraph. Using farms and planting plots as agent nodes, the edge weights are determined by the joint function of the potential gradient field of the absorption potential field distribution map and the traffic impedance of the logistics network. A multi-objective utility function is constructed, comprising logistics cost, agronomic suitability, environmental overload risk, and carbon emission cost. A potential field-game dual-channel attention mechanism is introduced to extract implicit supply and demand correlation features between nodes. The multi-objective utility function is dynamically weighted to generate a game utility matrix. Based on the Lagrange multiplier method, the environmental carrying capacity index is embedded as a hard constraint into the Nash equilibrium solution process. A differential game adaptive tracking algorithm is used to solve the batch allocation of manure and wastewater transfer and the multi-objective path optimization, outputting the globally optimal scheduling instruction set. The closed-loop feedback and adaptive correction module, connected to the intelligent matching and scheduling module and the field in-situ monitoring terminal, is used to receive soil available nutrient inversion data and crop growth indicators after manure application, construct observation residual sequences, compare the observation residual sequences with the predicted states of the physical information spatiotemporal graph neural network absorption potential field model in the corresponding grid and time window, calculate the joint posterior bias distribution through the online Bayesian update mechanism with shared hidden state space, and synchronously and inversely correct the decay dynamic parameters of the partial differential equation and the risk weighting coefficient of the multi-objective utility function, forming a closed-loop iterative control of evaluation-scheduling-feedback-correction.

2. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 1, characterized in that: The data acquisition and fusion module collects real-time data on manure output characteristics from the livestock farming sector, soil-crop background data from the planting sector, regional meteorological and hydrological data, and logistics network data. After spatiotemporal benchmark alignment and multimodal feature dimensionality reduction fusion, a standardized multidimensional feature tensor representing the correlation between planting and livestock elements is formed, specifically including: The data cleaning and standardization unit is used to access real-time data on manure output characteristics from livestock farms, soil-crop baseline data from planting farms, regional meteorological and hydrological data, and logistics network data. It identifies and interpolates outliers in daily manure removal volume, total nitrogen concentration, and total phosphorus concentration in the manure output characteristics data from livestock farms. It removes abnormal fluctuations in the nitrogen and phosphorus content of soil profiles and the nutrient demand curves during crop growth periods in the soil-crop baseline data from planting farms. It verifies the spatiotemporal consistency of rainfall and temperature sequences in the regional meteorological and hydrological data. It verifies the integrity of road passage times and restricted periods in the logistics network data, resulting in a cleaned, multi-source heterogeneous original dataset. The spatiotemporal reference alignment unit is used to receive the cleaned multi-source heterogeneous raw dataset. It aggregates the manure production trait data from the breeding end into a time series of equal length daily points according to the coordinates of the breeding farm center and the daily production rhythm. It maps the soil-crop background data from the planting end to a preset spatial discrete grid through hierarchical co-kriging interpolation and assigns multi-period soil nutrient and crop demand attributes to each grid. It downscales the regional meteorological and hydrological data to the same time resolution of the spatial discrete grid through spatiotemporal local weighted regression. It converts the logistics road network data into a time-varying road network impedance map with intersections as nodes and road segment traffic costs as edge weights. It unifies all data to the same coordinate reference system and timestamp index to form a spatiotemporally aligned multi-source dataset. The multimodal feature dimensionality reduction and fusion unit is used to construct a heterogeneous modal interactive Transformer network enhanced by position encoding, using spatiotemporally aligned multi-source datasets as input. The heterogeneous modal interactive Transformer network maps the characteristics of livestock manure release, soil absorption potential, meteorological and hydrological driving states, and road network accessibility features as modality-specific tokens. Through a multi-head cross-modal attention mechanism, it captures the supply and demand response relationship between manure supply and soil absorption, the regulatory relationship of meteorological and hydrological factors on nutrient decay, and the constraint pattern of road network impedance on flow paths. Under the information bottleneck variational autoencoder framework, it compresses high-dimensional modal interactive features into a low-dimensional latent space to generate a low-dimensional fused feature vector sequence. The tensor construction and output unit is used to receive the low-dimensional fused feature vector sequence, perform spatiotemporal partitioning and stacking of the low-dimensional fused feature vector sequence according to the grid number of the spatial discrete grid and the preset agricultural time window, and perform layer normalization processing on the stacked three-dimensional structure by feature channel to form a standardized multidimensional feature tensor representing the relationship between planting and breeding elements, and output the standardized multidimensional feature tensor to the dynamic absorption capacity assessment module.

3. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 1, characterized in that: The dynamic absorption capacity assessment module, based on standardized multidimensional feature tensors, constructs a physical information spatiotemporal graph neural network absorption potential field model, specifically including: The system receives the standardized multidimensional feature tensor output by the data acquisition and fusion module, extracts the geographic location identifier and grid adjacency relationship corresponding to the spatial discrete grid, takes each spatial discrete grid as a graph node, constructs directed edges representing the material flux between grids based on surface hydrological flow direction, topographic gradient and soil texture connectivity, and assigns a transmission attribute vector composed of soil permeability coefficient and nutrient diffusion coefficient to each directed edge, forming a directed graph topology that carries the initial structure of the absorption potential field. Based on the standardized multidimensional feature tensor, the nitrogen and phosphorus release potential energy of manure obtained by mapping the manure production trait data from the breeding end after spatiotemporal alignment in each spatial discrete grid, the existing soil nutrient reserves and expected absorption potential energy of the crop growth period obtained by parsing the soil-crop background data from the planting end, and the hydrothermal driving factors obtained by transforming regional meteorological and hydrological data are uniformly assigned as the initial node state vector of the corresponding graph node, so that the directed graph topology is accompanied by an initial state driven by multiple physical quantities. The partial differential equations describing the kinetics of nitrogen and phosphorus leaching, volatilization, and decay are discretized along the spatial dimension of the spatial discrete grid and the time dimension of the preset agricultural time window to construct a physical residual operator in the form of a difference. A differentiable physical regularization term is constructed with this physical residual operator as the core, and the differentiable physical regularization term is embedded into the loss function of the graph convolutional network to be trained to constrain the nutrient diffusion flux and decay process between adjacent grid nodes to conform to the laws of conservation of matter and kinetics. A potential inference network for nutrient absorption is constructed, consisting of multiple layers of heterogeneous graph convolutions. The network includes edge convolutional layers for transmitting the interaction between edge transmission attributes and node states, spatial graph attention layers for adaptively aggregating nutrient absorption features of neighboring nodes, and gated recurrent time coding units that introduce time-varying meteorological and hydrological driving sequences. The initial node state vector and transmission attribute vector are used as network inputs, and multi-scale nutrient absorption potential features are preserved through cross-layer skip connections, forming a graph neural network forward computation structure guided by physical regularization. By utilizing the standardized multidimensional feature tensors generated by the data acquisition and fusion module corresponding to the historical crop-livestock cycle period and the measured change sequence of soil nutrients in the field, a supervised training sample set for the absorption potential field is constructed. During training, the differentiable physical regularization term and the data fitting loss term between the predicted absorption potential energy value and the measured nutrient change are jointly optimized by backpropagation to obtain the weight parameters of the absorption potential field inference network, thus completing the construction of the physical information spatiotemporal graph neural network absorption potential field model.

4. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 3, characterized in that: The physical information spatiotemporal graph neural network absorption potential field model constructs a graph topology using spatial discrete grids as nodes and inter-grid material flux as edges. It captures the spatial diffusion relationship of nutrients through a graph convolutional network and embeds the residuals of the partial differential equations for nitrogen and phosphorus leaching and volatilization decay as differentiable regularization terms into the network loss function. Through spatiotemporal forward extrapolation, it calculates the nutrient supply and demand potential energy difference of each grid within a preset agricultural time window, outputting a spatiotemporal absorption potential field distribution map containing a dynamic absorption threshold and an environmental carrying capacity index. Specifically, it includes: The standardized multidimensional feature tensor received in real time is segmented and normalized according to the start time of the preset agricultural time window, and loaded into the physical information spatiotemporal graph neural network absorption potential field model that has been trained. Each spatial discrete grid node is assigned the initial value of nitrogen and phosphorus release potential energy of manure, the initial value of soil absorption background potential energy and hydrothermal driving state at the current time, thus completing the spatiotemporal initialization of the absorption potential field. Multi-step forward extrapolation is performed along the time axis with a time step within a preset agricultural time window. In each extrapolation step, the edge convolutional layer of the inference network calculates the nutrient flux between grids based on the transmission attribute vector and the node state at the previous time step. The spatial graph attention layer aggregates the neighbor influence and updates the node state. At the same time, the residual of the partial differential equation characterized by the differentiable physical regularization term is used to perform real-time constraint correction on the node state increment, so as to obtain the nitrogen and phosphorus mass stock, leaching loss and volatilization decay of each grid in the current step. After the simulation of each time step is completed, the existing soil nutrients in the node state vector are read and compared with the expected absorption amount of that step obtained by analyzing the crop growth period absorption curve. Combined with the correction coefficient of hydrothermal driving factor on nutrient availability, the difference between nutrient supply potential energy and absorption demand potential energy is calculated grid by grid to form the nutrient supply and demand potential energy difference of that step. The nutrient supply and demand potential energy difference of each grid is compared with the preset safety benchmark potential energy threshold based on soil environmental capacity. When the potential energy difference exceeds the safety benchmark potential energy threshold, the maximum amount of nitrogen and phosphorus equivalent of manure that the grid can absorb without exceeding the environmental risk threshold is marked as the dynamic absorption threshold. The ratio of the amount of manure already absorbed in the grid to the dynamic absorption threshold is converted into the current environmental carrying capacity index through nonlinear mapping. Collect the nutrient supply and demand potential energy difference, dynamic absorption threshold and environmental carrying capacity index of all grids at all time steps within the preset agricultural time window, perform co-kriging smoothing interpolation along the spatial dimension and temporal aggregation along the temporal dimension to generate a spatiotemporal absorption potential field distribution map with potential energy gradient representing the surplus or deficiency of absorption capacity and threshold boundary identifying the safe absorption space, and output the spatiotemporal absorption potential field distribution map to the intelligent matching and scheduling module.

5. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 1, characterized in that: The intelligent matching and scheduling module maps the spatiotemporal absorption potential field distribution map into a multi-agent scheduling hypergraph, specifically including: The system receives the spatiotemporal absorption potential field distribution map output by the dynamic absorption capacity assessment module, extracts the grids where the manure release potential energy in the spatial discrete grids covered by the grids is higher than the preset pollution generation start threshold as farm smart agent nodes, and extracts the grids where the planting plots with positive dynamic absorption thresholds and in the crop absorption period are located as planting plot smart agent nodes. Each smart agent node is assigned a node attribute vector consisting of potential energy value, current nutrient inventory status, geographical coordinates and function type label. Starting from the intelligent agent node of the breeding farm, streamline tracing is performed along the potential energy gradient vector field in the spatiotemporal absorption potential field distribution map. The intelligent agent node of the planting plot that the gradient streamline crosses and whose absorption demand potential energy exceeds the demand response threshold is included in the supply and demand attraction domain of the breeding farm. The intelligent agent node of the planting plot with multiple attraction domains is assigned soft membership degree according to the potential energy difference ratio of each upstream breeding farm node, forming a many-to-many supply and demand candidate relationship set. A scheduling hyperedge is defined as a unit for scheduling a batch of manure and sewage transfer. Each scheduling hyperedge connects a source node of a farm to at least one target node of a planting plot in a multi-to-multi supply and demand candidate relationship set. The total nitrogen and phosphorus equivalent load of the batch, the cumulative attenuation of the potential energy gradient along the path, and the time-varying shortest travel time extracted based on logistics network data are used as hyperedge attributes. All nodes, supply and demand candidate relationship edges and scheduling hyperedges are combined into a multi-agent scheduling hypergraph.

6. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 5, characterized in that: Using farms and planting plots as intelligent agent nodes, and determining edge weights using a joint function of the potential energy gradient field of the absorption potential field distribution map and the traffic impedance of the logistics road network, a multi-objective utility function is constructed with components including logistics cost, agronomic suitability, environmental overload risk, and carbon emission cost. Specifically, this includes: For each pair of candidate connections between farm agent nodes and planting plot agent nodes in the multi-agent scheduling hypergraph, the potential energy gradient vector sequence on the corresponding transportation path of the candidate connection is extracted from the Canna potential field distribution map. At the same time, the time-varying traffic impedance sequence of the same path is extracted from the logistics network data. The projection modulus of the potential energy gradient vector on the tangential direction of the path is used as the nutrient transport compliance component, and the equivalent value of the traffic impedance after time cost is used as the transport obstacle component. The entropy weight method is used to fuse the compliance component and the obstacle component to obtain the path joint weight. The environmental carrying capacity index attenuation penalty term of the passing grid is embedded in the path joint weight to form the edge weight that characterizes the comprehensive cost of manure transfer. Logistics cost is defined as the product of edge weight and unit distance transportation cost rate; agronomic fit is defined as the sliding convolution matching degree of the manure nutrient release time curve and the crop growth period nutrient absorption time curve of the planting plot; environmental overload risk is defined as the conditional risk probability that the nitrogen and phosphorus stock of the planting plot exceeds the dynamic absorption threshold after receiving manure allocation; and carbon emission cost is defined as the product of the carbon emission factor of the transport vehicle and the mileage driven by the load. The logistics cost, agronomic fit, environmental overload risk and carbon emission cost are respectively processed by range standardization, and then the four types of standardized loss values ​​are synthesized into a multi-objective utility function with the manure allocation amount of each candidate connection as the independent variable through a weighted Chebyshev aggregation function.

7. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 6, characterized in that: A potential-game dual-channel attention mechanism is introduced to extract implicit supply and demand correlation features between nodes. The multi-objective utility function is dynamically weighted to generate a game utility matrix. Based on the Lagrange multiplier method, the environmental carrying capacity index is embedded as a hard constraint into the Nash equilibrium solution process. A differential game adaptive tracking algorithm is used to solve the batch allocation and multi-objective path optimization for sewage transfer, outputting a globally optimal scheduling instruction set, specifically including: The node attribute vectors and hyperedge attributes in the multi-agent scheduling hypergraph are encoded as query vectors, key vectors and value vectors, and the relative position encoding derived from the potential energy gradient field is injected. The nutrient supply and demand potential energy complementarity between the aquaculture farm node and the planting plot node is calculated by scaling dot product attention, and a potential field attention weight matrix is ​​generated to highlight the supply and demand pairing with large potential energy drop and low transmission loss. Using the agents of each farm and the agents of each plot of land as game participants, the historical strategy payoff time difference sequence of each participant in the previous scheduling rolling window is extracted. The implicit competition-cooperation time dependence among the participants in the non-cooperative elimination game is learned by using the causal mask multi-head self-attention mechanism, and the game attention weight matrix is ​​generated to express the priority elimination tendency of each participant towards other participants. The potential field attention weight matrix and the game attention weight matrix are fused by element-wise multiplication and normalized to obtain a dynamic weighted coefficient matrix. The dynamic weighted coefficient matrix is ​​used to adaptively scale the utility components of each farm-plot pair in the multi-objective utility function to generate a game utility matrix. The elements of this game utility matrix represent the comprehensive utility of unit manure equivalent transfer under the condition of simultaneously reflecting the physical absorption law and the multi-subject strategy preference. Using the environmental carrying capacity index of all planting plots at each time point not being lower than the preset environmental safety lower limit as a rigid constraint, the Lagrange multiplier vector corresponding to the plot constraint is introduced to transform the constrained multi-agent scheduling optimization problem into a generalized Nash equilibrium problem in the form of an augmented Lagrange function. The constraint violation quantity is incorporated into the local utility function of each agent in the form of a penalty term. Based on the first-order optimality condition of the generalized Nash equilibrium problem, a system of differential algebraic equations is constructed. A differential game adaptive tracking algorithm is adopted, which uses the current scheduling state as the initial value and performs numerical integration along the direction of the synthesis of utility gradient and constraint gradient to track the equilibrium trajectory. At each time step, the batch allocation of manure, the flow of the transportation path and the Lagrange multiplier are updated synchronously until the allocation state converges to the global Nash equilibrium point. Finally, the converged farm batch loading, target plot, transportation path node sequence, departure time window and segment allocation are compiled into a global optimal scheduling instruction set and output.

8. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 1, characterized in that: The closed-loop feedback and adaptive correction module receives soil available nutrient inversion data and crop growth indicators after manure application, and constructs an observation residual sequence, specifically including: The system receives real-time data from in-situ field monitoring terminals, such as soil ion-selective electrodes, near-infrared spectral probes, and multispectral canopy imagers, which are deployed in the planting area. The data includes soil available nitrogen content, soil available phosphorus content, soil moisture content, soil temperature, crop leaf area index, and normalized vegetation index. The system performs bilinear spatiotemporal matching based on timestamps and spatial discrete grids to generate a gridded field observation snapshot sequence. Anomaly detection based on adaptive threshold and local weighted scatter smoothing and filling are performed on the gridded field observation snapshot sequence. Then, the soil available nutrient sequence and crop growth index sequence are recursively estimated using time-domain Kalman filtering, and the smoothed soil available nutrient status estimation sequence and crop growth status estimation sequence are output. The pre-set crop nutrient uptake inversion model is called to convert the smoothed crop growth status estimation sequence into the actual crop nitrogen and phosphorus uptake flux at each time step. Based on the smoothed soil available nutrient status estimation sequence and the actual crop nitrogen and phosphorus uptake flux, the measured soil nutrient net storage variable sequence is obtained through grid mass conservation accounting. Based on the baseline value of soil nutrients in the grid before the execution of the scheduling command and the theoretical nutrient input contained in the batch allocation plan of manure and sewage output by the intelligent matching scheduling module, an interference-free reference absorption trajectory is constructed. The deviation between the measured net soil nutrient storage variable sequence and the interference-free reference absorption trajectory is calculated step by step to generate an observation residual sequence that characterizes the deviation between the actual absorption effect and the planned absorption target.

9. The livestock and poultry manure disposal capacity assessment and intelligent matching scheduling system for integrated crop and livestock farming as described in claim 8, characterized in that: The observed residual sequence is compared with the predicted states of the physical information spatiotemporal graph neural network absorption potential field model at the corresponding grid and time window. The joint posterior bias distribution is calculated through an online Bayesian update mechanism sharing the hidden state space. Based on this, the decay dynamic parameters of the partial differential equation and the risk weighting coefficients of the multi-objective utility function are synchronously and inversely corrected, forming a closed-loop iterative control of evaluation-scheduling-feedback-correction, specifically including: The spatiotemporal graph neural network model for extracting physical information, within the same spatial discrete grid and the same feedback time window corresponding to the observed residual sequence, generates predicted soil nutrient stock, predicted nitrogen and phosphorus leaching flux, and predicted volatilization attenuation flux sequences through spatiotemporal forward extrapolation calculations. These sequences serve as the model's predicted state. The observed residual sequence is time-stepped to align with the model's predicted state, and the observed residual sequence is used to correct and restore the model's predicted state, thus constructing a normalized information sequence that reflects the model's prediction bias. The normalized information sequence is injected into a shared hidden state encoder consisting of long short-term memory units and graph attention layers. Under the premise of preserving the original hidden state space topology of the physical information spatiotemporal graph neural network absorption potential field model, the encoder extracts the first hidden state vector representing the deviation of the decay dynamic parameters of the partial differential equation, and the second hidden state vector representing the deviation of the weighted coefficient of environmental overload risk in the multi-objective utility function. Using the first and second hidden state vectors as observed variables, and employing the particle filter online Bayesian recursive framework, a set of weighted random particles is maintained in the decay dynamics parameter space and the risk weighting coefficient space, respectively. Through importance resampling and sequential updates, the joint posterior probability distribution of the decay dynamics parameters and the joint posterior probability distribution of the risk weighting coefficients are recursively calculated, and the expected values ​​of the two are used as the maximum a posteriori estimate of the joint posterior bias distribution. Based on the expected distribution of the joint posterior bias of the decay kinetic parameters, the microbial decomposition rate constant, ammonia volatilization coefficient, and denitrification rate coefficient in the partial differential equations for nitrogen and phosphorus leaching and volatilization decay are adjusted in reverse. The adjusted parameters are then written into the calculation graph of the differentiable physical regularization term of the physical information spatiotemporal graph neural network absorption potential field model, and the model parameters of the dynamic absorption capacity assessment module are updated synchronously. At the same time, based on the expected distribution of the joint posterior bias of the risk weighting coefficients, the weighting factors corresponding to the environmental overload risk component of the multi-objective utility function in the intelligent matching and scheduling module are corrected. After the parameters are corrected, the absorption potential field is re-evaluated using the updated physical information spatiotemporal graph neural network absorption potential field model, and the next round of intelligent matching scheduling is triggered, forming a closed-loop iterative control loop from evaluation, scheduling, field execution, feedback monitoring to model correction.

Images