Intelligent farmland irrigation decision-making method based on digital twinning

Abstract

The invention discloses an intelligent farmland irrigation decision-making method based on digital twinning, and the method comprises the following steps: S1, collecting multi-source data of a farmland environment, and generating a time sequence farmland environment data set; s2, constructing a farmland digital twinborn model, and updating node and edge attribute information in real time; s3, constructing a time sequence diagram based on the updated twinborn model, and inputting the time sequence diagram to a time sequence diagram neural network; s4, optimizing structural parameters of the neural network of the time sequence diagram by applying a red-mouth blue-magpie optimization algorithm; s5, predicting the future soil water potential and the crop water demand trend by using the optimized network; s6, generating an irrigation control instruction according to the prediction result in combination with water resource supply; and S7, executing irrigation, monitoring the humidity of the land parcel and the crop state in real time, and updating the farmland digital twinborn model. According to the invention, by fusing the digital twinning, the time sequence diagram neural network and the Red-billed-blue-magpie optimization algorithm, accurate prediction and intelligent irrigation control of the future water demand trend of the farmland are realized, and the water resource utilization efficiency and the crop yield are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of geographic information technology, and in particular to an intelligent farmland irrigation decision-making method based on digital twins. Background Art

[0002] As global agricultural production shifts toward intelligent and refined production, farmland irrigation, a crucial component of the agricultural planting process, has garnered widespread attention for its intelligent control and optimization decisions. Traditional farmland irrigation methods rely primarily on manual experience or simple timing control systems. These methods are unable to perceive real-time changes in the farmland microenvironment and struggle to dynamically adjust according to actual crop water requirements. This leads to significant water waste, low irrigation efficiency, and the susceptibility to crop water stress and yield decline. To improve the scientific nature and precision of farmland irrigation, academia and industry have proposed intelligent irrigation systems based on environmental sensor data, Internet of Things (IoT) technology, and big data analysis.

[0003] Existing intelligent irrigation systems typically rely on static models or rule engines, collecting environmental data through soil moisture sensors and meteorological monitoring equipment. These systems then make irrigation decisions based on preset threshold rules or simple machine learning prediction models. While these systems have improved automation to a certain extent, they still face numerous challenges: First, data processing and decision-making models are often static or low-dynamic, making them ineffective in addressing the rapidly changing nature of farmland environments over time. Second, these models lack spatial modeling capabilities, often making local judgments based on single-point data points and lacking modeling of spatial characteristics such as water migration between different plots, topographic relationships, and crop interactions. Third, in terms of temporal modeling, these systems lack long-term trend predictions for soil water potential evolution and crop dynamic water demand patterns, relying solely on short-term or single-step predictions, which can lead to delayed or overly predictive irrigation decisions. Fourth, they lack system integration for dynamically constrained water supply conditions, making it prone to policy failures when water supply becomes tight.

[0004] In recent years, graph neural networks (GNNs) have been introduced into the field of intelligent farmland decision-making due to their powerful structured data processing capabilities. They can, to a certain extent, model the spatial relationships between plots and enhance the ability to couple local and global analysis. At the same time, time series prediction technologies, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), have demonstrated promising results in crop growth forecasting and water resource demand analysis. However, traditional GNN or time series prediction methods often use independent modeling, lacking a close integration of spatial and temporal features and an effective joint spatiotemporal modeling mechanism. This makes it difficult to balance spatial dependencies and temporal evolution, limiting the intelligence level of overall irrigation decision-making systems.

[0005] To address these issues, some studies have proposed modeling approaches based on temporal graph neural networks (TGNNs), attempting to introduce dynamic changes in the temporal dimension into the graph structure, taking into account the temporal evolution of spatial relationships between plots and environmental conditions. However, existing TGNN methods still have the following shortcomings: First, in the spatiotemporal feature fusion stage, simple splicing or weighting methods are often used, which cannot flexibly and dynamically adjust the spatiotemporal dependency ratio according to the actual changing farmland environment. Second, when modeling the graph structure, there is a lack of a mechanism to dynamically intervene in the relationships between plots due to external environmental factors (such as rainfall, wind speed, and temperature anomalies), resulting in poor adaptability to sudden changes. Third, in terms of temporal encoding, fixed or static position encoding methods are often used, which fails to fully utilize the periodic laws in the farmland environment (such as day and night alternation and seasonal changes), resulting in insufficient temporal feature extraction.

[0006] Furthermore, traditional methods for training and optimizing time-series graph neural networks often rely on manual hyperparameter adjustment or conventional grid search and random search techniques. These methods are inefficient and prone to falling into local optima, making it difficult to find the optimal model parameter combination that adapts to the dynamic characteristics of farmland environments. Furthermore, existing intelligent irrigation systems often fail to fully integrate digital twin technology into their frameworks, remaining limited to sensor data collection and simple prediction. They lack dynamic synchronization and feedback correction mechanisms for real-time data based on farmland twins, making it impossible to achieve close mapping and coordinated optimization between the physical and virtual worlds of farmland, severely limiting the system's adaptive capabilities and intelligent decision-making capabilities.

[0007] Therefore, how to provide an intelligent farmland irrigation decision-making method based on digital twins is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose an intelligent farmland irrigation decision-making method based on digital twins. The present invention makes full use of the parameter optimization mechanism of farmland digital twin modeling, dynamic prediction of time series graph neural network and red-billed blue magpie optimization algorithm, and describes in detail the whole process of realizing the prediction of future water demand trends of farmland plots and the generation of intelligent irrigation control instructions. It has the advantages of high prediction accuracy, high water resource utilization rate, high level of decision-making intelligence and strong system adaptability.

[0009] According to an embodiment of the present invention, a smart farmland irrigation decision-making method based on digital twins includes the following steps:

[0010] S1. Collect multi-source data of farmland environment, pre-process the multi-source data, and generate a time series farmland environment dataset;

[0011] S2. Based on the time-series farmland environmental dataset, a farmland digital twin model is constructed. The real-time collected environmental data is dynamically input into the farmland digital twin model to update the attribute information of nodes and edges.

[0012] S3. Based on the updated farmland digital twin model, a time series graph is constructed, time series changes are encoded into a graph structure, and input into a time series graph neural network for modeling. The structural parameters and training hyperparameters of the time series graph neural network are preliminarily set.

[0013] S4. Apply the red-billed blue magpie optimization algorithm to optimize the structural parameters and training hyperparameters of the time series graph neural network to obtain the optimal parameter configuration of the time series graph neural network;

[0014] S5. Using the optimized time series graph neural network, the soil water potential evolution trend and crop water demand trend of each farmland plot in multiple future time steps are predicted, and the predicted water demand value of each plot corresponding to the time step is output;

[0015] S6. Based on the prediction results and combined with the farmland water resources supply conditions, generate farmland irrigation control instructions, including the irrigation start time, irrigation duration, irrigation water volume and irrigation sequence for each plot;

[0016] S7. During the irrigation process, the soil moisture changes and crop status changes of each plot are monitored in real time, feedback data is collected, and the feedback data is input into the farmland digital twin model. The attribute values ​​of the nodes and edges are corrected in real time, the twin status is updated, and the corrected farmland digital twin model is used for the prediction and optimization cycle of the next cycle.

[0017] Optionally, the multi-source data specifically includes soil moisture data, meteorological data, crop growth status data and plot geographic location information, which are used to construct a digital twin model of farmland and support dynamic modeling and prediction of time series graph neural networks.

[0018] Optionally, the preprocessing of multi-source data specifically includes data cleaning, missing value filling, outlier removal and normalization processing, so as to improve the quality of multi-source data.

[0019] Optionally, the S2 specifically includes:

[0020] S21. Based on the time series farmland environment dataset, the farmland plot set is recorded as the node set V = {v1, v2, ..., v n}, the physical connection between farmland plots is recorded as edge set E = {e ij |v i ,v j ∈V}, construct the initial farmland graph G=(V,E), where v irepresents the i-th farmland plot, i∈{1,2,…,n}, n is the total number of farmland plot nodes, e ij Represents node v i With node v j There is a direct physical connection between

[0021] S22. Define each node v i The eigenvector x i =[s i ,t i ,c i ], where s i For plot v i Soil moisture value, t i For plot v i The current temperature value, c i For plot v i Corresponding crop growth status index value;

[0022] S23. Define each edge e ij The eigenvalue w ij , representing the plot v i With plot v j The water flow trend between ij Obtained through comprehensive calculation based on the height difference, soil permeability and distance between plots;

[0023] S24, the farmland environment data collected in real time is streamed Input into the initial farmland map G, where Represents plot v i At time t, the observation data, i∈{1,2,…,n}, updates the node v i The eigenvector of and edge ij The eigenvalue of

[0024] S25, through the node feature vector and edge eigenvalues Continuous dynamic update to form a real-time evolving farmland digital twin model G t =(V,E,X t ,W t ),in

[0025] S26, update the farmland digital twin model G t Serves as the basis for timing graph construction and timing graph neural network input.

[0026] Optionally, the S3 specifically includes:

[0027] S31. In the updated farmland digital twin model G t Based on this, we define the node set V, edge set E, and node feature matrix X t , edge weight matrix W t , construct a time series diagram with time series T

[0028] S32, at each time step t k On the top, extract node v i Features and edge weights Graph convolution operations are used to extract spatial features, and time recursive units are used to model temporal characteristics;

[0029] S33, introduce the spatiotemporal dual attention mechanism, perform weighted fusion of spatial features and temporal features, and set the spatiotemporal interaction weight modulation coefficient λ st , the fusion feature is expressed as:

[0030]

[0031] in, Represents node v i At time step t k The fused spatiotemporal joint feature representation vector, For node v i At time step t k The extracted spatial feature representation, It is the time feature representation;

[0032] S34, in the process of graph structure construction, based on the external factor dataset F t Dynamically adjust edge weights and define external sensitivity coefficient β ext , updated edge weights for:

[0033]

[0034] in, Indicates the rate of change of external factors on plot v i With v j The impact value of the relationship;

[0035] S35. In the time coding process, a cycle-aware time coding method is used to encode each time step t. k Generate periodic encoding vector And introduce the period weight coefficient γ per , and obtain the improved temporal feature representation:

[0036]

[0037] in, For the tth k The time encoding vector after the fusion of time steps, is the absolute time code, Encode the cycle time;

[0038] S36. Preliminary setting of structural parameters of the temporal graph neural network, including the number of graph convolution layers L g , time step expansion number m, node feature dimension d v , the number of hidden layer neurons h, and the newly added spatiotemporal interaction weight modulation coefficient λ st , external sensitivity coefficient β ext , period weight coefficient γ per , and set the training hyperparameters, including learning rate α, batch size b, and optimizer parameter set θ, to complete the initialization modeling of the timing graph neural network.

[0039] Optionally, the S4 specifically includes:

[0040] S41, modulating the coefficient λ with the spatiotemporal interaction weight st , external sensitivity coefficient β ext , period weight coefficient γ per As the optimization object, construct the parameter set to be optimized Θ new ={λ st ,β ext ,γ per};

[0041] S42, using a population generation method based on a segmented initialization strategy to initialize the red-billed blue magpie optimization algorithm population, each red-billed blue magpie individual Θ new,p Randomly select initial values ​​from the divided reasonable parameter intervals, where p is the individual number of the red-billed blue magpie in the population;

[0042] S43. Define comprehensive fitness function Using parameter Θ new,p The prediction error of the trained time series graph neural network on the validation set is weighted with the water resource utilization efficiency as the evaluation criteria:

[0043]

[0044] in, Configure Θ for parameters new,p The predicted output of the next i-th sample, y i is the true label of sample i, N is the number of samples in the validation set, η(Θ new,p ) is the water resource utilization efficiency corresponding to the farmland irrigation strategy derived from the prediction results, ω1 and ω2 are weighting coefficients;

[0045] S44. According to the foraging behavior rules of the red-billed blue magpie, the individual Θ of the red-billed blue magpie population new,p To perform a location update:

[0046]

[0047] in, is the red-billed blue magpie individual with the best fitness in the tth generation, is another randomly selected red-billed blue magpie individual, r1 and r2 are random numbers in the range of [0,1]. is the updated parameter configuration vector of the red-billed blue magpie individual in the p-th population of the t+1 generation, is the updated parameter configuration vector of the red-billed blue magpie individual in the p-th population of the t-th generation;

[0048] S45. Introducing a dynamic foraging radius adjustment mechanism to dynamically scale the search step size based on the fitness ranking of individual blue magpies in the population, expanding the exploration range in the early stages of the search and reducing the step size in the later convergence stages, thereby improving the global and local convergence of the search.

[0049] S46. Introduce an individual variation mechanism based on multiple foraging strategies. When there is no fitness improvement for several consecutive generations, randomly adjust the single parameter λ of some low-fitness red-billed blue magpie individuals. st , β ext or γ per Perform disturbance mutation to prevent the population from falling into local optimality;

[0050] S47, when the maximum number of iterations T is reached max Or after the fitness converges, the optimization process is terminated and the optimal parameter configuration is output The optimal parameter configuration is applied to the time series graph neural network for predictive modeling of soil water potential and crop water demand trends in farmland plots.

[0051] Optionally, the S5 specifically includes:

[0052] S51. Based on the real-time updated farmland digital twin model, extract the node feature information of each plot at the current moment and the edge weight information between plots;

[0053] S52, using the optimized temporal graph neural network, inputting node feature information and the relationship between plots, and performing feature encoding using the optimized structural parameters and the newly added fusion mechanism;

[0054] S53, based on the input current time data, recursively predict the state changes of each farmland plot in multiple future time steps through the optimized time series graph neural network;

[0055] S54. At each future time step, output the soil water potential evolution trend prediction results corresponding to each block, and record the dynamic changes of the soil water status of each block over time;

[0056] S55, based on the predicted soil water potential trend, further derive and output crop water requirement trend data for each plot at each future time step;

[0057] S56. Arrange the predicted water demand values ​​corresponding to each block at multiple future time steps to generate a farmland water demand prediction dataset.

[0058] Optionally, the S6 specifically includes:

[0059] S61. Based on the future water demand prediction results output by the optimized time series graph neural network, extract the water demand trend data of each farmland plot at each future time step;

[0060] S62. Collect data on water supply conditions in the farmland area in the current and foreseeable time period, including water supply volume, pump station operating status, and channel delivery capacity;

[0061] S63. Based on water demand trend data and water resource supply conditions, conduct a water demand matching analysis at the plot level to determine a feasible irrigation start time period and preliminary water allocation plan;

[0062] S64. Develop irrigation priorities for each plot based on the urgency of water demand and water supply capacity limitations, and generate an irrigation ranking list;

[0063] S65. Determine the specific irrigation start time, irrigation duration, and irrigation water volume for each plot according to the irrigation ranking list, and form a specific plot irrigation control instruction set;

[0064] S66. Integrate the irrigation control instructions of each plot to generate a global farmland irrigation control plan and execute the scheduling operation.

[0065] The beneficial effects of the present invention are:

[0066] By introducing digital twin technology, the present invention constructs a real-time, dynamic, and synchronized virtual twin of farmland based on multi-source data of the farmland environment. This enables real-time mapping and feedback updates of the states between the physical farmland and the virtual system. Compared to traditional irrigation systems that rely solely on single static sensor data collection, this system can continuously perceive changes in the farmland environment, enhancing the real-time and accuracy of the system. By constructing a time-series graph neural network based on the farmland digital twin model, it comprehensively considers the spatial relationship between plots and the dynamic characteristics of environmental states over time, effectively extracting soil water potential evolution and crop water demand trends, and improving the model's stability and accuracy in multi-time-step continuous prediction tasks. In response to the shortcomings of traditional time series modeling methods in spatiotemporal feature fusion, external environmental intervention modeling, and periodic time feature processing, the present invention innovatively introduces a spatiotemporal interaction weight modulation mechanism, an external sensitivity dynamic adjustment mechanism, and a periodic perception time coding mechanism, which can flexibly adapt to the complex dynamic changes of the farmland environment and improve the model's adaptability to extreme meteorological conditions and seasonal fluctuations.

[0067] The present invention further utilizes the red-billed blue magpie optimization algorithm to dynamically optimize the newly added key structural parameters in the time-series graph neural network, avoiding the low efficiency and easy trapping of local optimality of traditional hyperparameter adjustment methods. It can dynamically obtain the optimal network configuration according to different farmland environments and crop growth cycle changes, effectively improving the overall prediction performance and resource allocation intelligence level. Through the optimized time-series graph neural network, the present invention can accurately predict the soil water potential change trend and crop water demand change trend of each farmland plot in multiple time steps in the future. On this basis, combined with the farmland water resource supply conditions, it dynamically generates precise control instructions including irrigation start time, duration, water volume and irrigation sequence, realizing accurate, intelligent and efficient farmland irrigation scheduling. Compared with existing intelligent irrigation systems, the present invention significantly improves water resource utilization efficiency, reduces the risk of over-irrigation or under-irrigation, reduces manual intervention operations, improves the automation and intelligence level of agricultural production, and has good scalability and applicability, and can adapt to the intelligent irrigation needs of different scales and different types of crop planting environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is a flow chart of an intelligent farmland irrigation decision-making method based on digital twins proposed in this invention;

[0070] Figure 2This is a structural schematic diagram of the farmland digital twin model construction and dynamic update process of the digital twin-based intelligent farmland irrigation decision-making method proposed in the present invention. DETAILED DESCRIPTION

[0071] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0072] refer to Figure 1 and Figure 2 , an intelligent farmland irrigation decision-making method based on digital twins, including the following steps:

[0073] S1. Collect multi-source data of farmland environment, pre-process the multi-source data, and generate a time series farmland environment dataset;

[0074] S2. Based on the time-series farmland environmental dataset, a farmland digital twin model is constructed. The real-time collected environmental data is dynamically input into the farmland digital twin model to update the attribute information of nodes and edges.

[0075] S3. Based on the updated farmland digital twin model, a time series graph is constructed, time series changes are encoded into a graph structure, and input into a time series graph neural network for modeling. The structural parameters and training hyperparameters of the time series graph neural network are preliminarily set.

[0076] S4. Apply the red-billed blue magpie optimization algorithm to optimize the structural parameters and training hyperparameters of the time series graph neural network to obtain the optimal parameter configuration of the time series graph neural network;

[0077] S5. Using the optimized time series graph neural network, the soil water potential evolution trend and crop water demand trend of each farmland plot in multiple future time steps are predicted, and the predicted water demand value of each plot corresponding to the time step is output;

[0078] S6. Based on the prediction results and combined with the farmland water resources supply conditions, generate farmland irrigation control instructions, including the irrigation start time, irrigation duration, irrigation water volume and irrigation sequence for each plot;

[0079] S7. During the irrigation process, the soil moisture changes and crop status changes of each plot are monitored in real time, feedback data is collected, and the feedback data is input into the farmland digital twin model. The attribute values ​​of the nodes and edges are corrected in real time, the twin status is updated, and the corrected farmland digital twin model is used for the prediction and optimization cycle of the next cycle.

[0080] In this embodiment, the multi-source data specifically includes soil moisture data, meteorological data, crop growth status data and plot geographic location information, which are used to construct a digital twin model of farmland and support dynamic modeling and prediction of time series graph neural networks.

[0081] In this embodiment, the preprocessing of multi-source data specifically includes data cleaning, missing value filling, outlier removal and normalization processing, which are used to improve the quality of multi-source data.

[0082] In this embodiment, S2 specifically includes:

[0083] S21. Based on the time series farmland environment dataset, the farmland plot set is recorded as the node set V = {v1, v2, ..., v n}, the physical connection between farmland plots is recorded as edge set E = {e ij |v i ,v j ∈V}, construct the initial farmland graph G=(V,E), where v i represents the i-th farmland plot, i∈{1,2,…,n}, n is the total number of farmland plot nodes, e ij Represents node v i With node v j There is a direct physical connection between

[0084] S22. Define each node v i The eigenvector x i =[s i ,t i ,c i ], where s i For plot v i Soil moisture value, t i For plot v i The current temperature value, c i For plot v i Corresponding crop growth status index value;

[0085] S23. Define each edge e ij The eigenvalue w ij , representing the plot v i With plot v j The water flow trend between ij Obtained through comprehensive calculation based on the height difference, soil permeability and distance between plots;

[0086] S24, the farmland environment data collected in real time is streamed Input into the initial farmland map G, where Represents plot v iAt time t, the observation data, i∈{1,2,…,n}, updates the node v i The eigenvector of and edge ij The eigenvalue of

[0087] S25, through the node feature vector and edge eigenvalues Continuous dynamic update to form a real-time evolving farmland digital twin model G t =(V,E,X t ,W t ),in

[0088] S26, update the farmland digital twin model G t Serves as the basis for timing graph construction and timing graph neural network input.

[0089] In this embodiment, S3 specifically includes:

[0090] S31. In the updated farmland digital twin model G t Based on this, we define the node set V, edge set E, and node feature matrix X t , edge weight matrix W t , construct a time series diagram with time series T

[0091] S32, at each time step t k On the top, extract node v i Features and edge weights Graph convolution operations are used to extract spatial features, and time recursive units are used to model temporal characteristics;

[0092] S33, introduce the spatiotemporal dual attention mechanism, perform weighted fusion of spatial features and temporal features, and set the spatiotemporal interaction weight modulation coefficient λ st , the fusion feature is expressed as:

[0093]

[0094] in, Represents node v i At time step t k The fused spatiotemporal joint feature representation vector, For node v i At time step t k The extracted spatial feature representation, It is the time feature representation;

[0095] S34, in the process of graph structure construction, based on the external factor dataset Ft Dynamically adjust edge weights and define external sensitivity coefficient β ext , updated edge weights for:

[0096]

[0097] in, Indicates the rate of change of external factors on plot v i With v j The impact value of the relationship;

[0098] S35. In the time coding process, a cycle-aware time coding method is used to encode each time step t. k Generate periodic encoding vector And introduce the period weight coefficient γ per , and obtain the improved temporal feature representation:

[0099]

[0100] in, For the tth k The time encoding vector after the fusion of time steps, is the absolute time code, Encode the cycle time;

[0101] S36. Preliminary setting of structural parameters of the temporal graph neural network, including the number of graph convolution layers L g , time step expansion number m, node feature dimension d v , the number of hidden layer neurons h, and the newly added spatiotemporal interaction weight modulation coefficient λ st , external sensitivity coefficient β ext , period weight coefficient γ per , and set the training hyperparameters, including learning rate α, batch size b, and optimizer parameter set θ, to complete the initialization modeling of the timing graph neural network.

[0102] In this embodiment, the S4 specifically includes:

[0103] S41, modulating the coefficient λ with the spatiotemporal interaction weight st , external sensitivity coefficient β ext , period weight coefficient γ per As the optimization object, construct the parameter set to be optimized Θ new ={λ st ,β ext ,γ per};

[0104] S42, using a population generation method based on a segmented initialization strategy to initialize the red-billed blue magpie optimization algorithm population, each red-billed blue magpie individual Θnew,p Randomly select initial values ​​from the divided reasonable parameter intervals, where p is the individual number of the red-billed blue magpie in the population;

[0105] S43. Define comprehensive fitness function Using parameter Θ new,p The prediction error of the trained time series graph neural network on the validation set is weighted with the water resource utilization efficiency as the evaluation criteria:

[0106]

[0107] in, Configure Θ for parameters new,p The predicted output of the next i-th sample, y i is the true label of sample i, N is the number of samples in the validation set, η(Θ new,p ) is the water resource utilization efficiency corresponding to the farmland irrigation strategy derived from the prediction results, ω1 and ω2 are weighting coefficients;

[0108] S44. According to the foraging behavior rules of the red-billed blue magpie, the individual Θ of the red-billed blue magpie population new,p To perform a location update:

[0109]

[0110] in, is the red-billed blue magpie individual with the best fitness in the tth generation, is another randomly selected red-billed blue magpie individual, r1 and r2 are random numbers in the range of [0,1]. is the updated parameter configuration vector of the red-billed blue magpie individual in the p-th population of the t+1 generation, is the updated parameter configuration vector of the red-billed blue magpie individual in the p-th population of the t-th generation;

[0111] S45. Introducing a dynamic foraging radius adjustment mechanism to dynamically scale the search step size based on the fitness ranking of individual blue magpies in the population, expanding the exploration range in the early stages of the search and reducing the step size in the later convergence stages, thereby improving the global and local convergence of the search.

[0112] S46. Introduce an individual variation mechanism based on multiple foraging strategies. When there is no fitness improvement for several consecutive generations, randomly adjust the single parameter λ of some low-fitness red-billed blue magpie individuals. st , β ext or γ per Perform disturbance mutation to prevent the population from falling into local optimality;

[0113] S47, when the maximum number of iterations T is reached max Or after the fitness converges, the optimization process is terminated and the optimal parameter configuration is output The optimal parameter configuration is applied to the time series graph neural network for predictive modeling of soil water potential and crop water demand trends in farmland plots.

[0114] In this embodiment, the S5 specifically includes:

[0115] S51. Based on the real-time updated farmland digital twin model, extract the node feature information of each plot at the current moment and the edge weight information between plots;

[0116] S52, using the optimized temporal graph neural network, inputting node feature information and the relationship between plots, and performing feature encoding using the optimized structural parameters and the newly added fusion mechanism;

[0117] S53, based on the input current time data, recursively predict the state changes of each farmland plot in multiple future time steps through the optimized time series graph neural network;

[0118] S54. At each future time step, output the soil water potential evolution trend prediction results corresponding to each block, and record the dynamic changes of the soil water status of each block over time;

[0119] S55, based on the predicted soil water potential trend, further derive and output crop water requirement trend data for each plot at each future time step;

[0120] S56. Arrange the predicted water demand values ​​corresponding to each block at multiple future time steps to generate a farmland water demand prediction dataset.

[0121] In this embodiment, S6 specifically includes:

[0122] S61. Based on the future water demand prediction results output by the optimized time series graph neural network, extract the water demand trend data of each farmland plot at each future time step;

[0123] S62. Collect data on water supply conditions in the farmland area in the current and foreseeable time period, including water supply volume, pump station operating status, and channel delivery capacity;

[0124] S63. Based on water demand trend data and water resource supply conditions, conduct a water demand matching analysis at the plot level to determine a feasible irrigation start time period and preliminary water allocation plan;

[0125] S64. Develop irrigation priorities for each plot based on the urgency of water demand and water supply capacity limitations, and generate an irrigation ranking list;

[0126] S65. Determine the specific irrigation start time, irrigation duration, and irrigation water volume for each plot according to the irrigation ranking list, and form a specific plot irrigation control instruction set;

[0127] S66. Integrate the irrigation control instructions of each block to generate a global farmland irrigation control plan and execute the scheduling operation.

[0128] Example 1:

[0129] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a winter wheat planting base, which covers an area of ​​about 60 mu and mainly grows winter wheat. The plots are carefully divided and the water supply mode is rotated. The traditional irrigation method relies heavily on manual experience and has a series of problems such as uneven irrigation, waste of water resources, and inconsistent crop growth. Especially in the critical period of wheat jointing in late winter and early spring, the soil water potential changes rapidly, but the response of artificial irrigation is delayed, which often leads to water stress in crops and affects the final yield and quality. The present invention aims to solve the problems of insufficient perception of soil water potential changes by traditional irrigation systems, slow response to crop water demand trends, and low efficiency of water resource allocation, so as to achieve accurate prediction of future water demand status of farmland and intelligent irrigation control.

[0130] In the actual deployment, environmental monitoring equipment for soil moisture, soil temperature, evaporation, and meteorological data was first installed on the farmland. Drone aerial surveys were then used to segment and model the plots. A digital twin model of the farmland was constructed, using plots as basic nodes and terrain elevation differences, soil permeability, and drainage channel relationships as edge weights. All sensor data and data on external environmental changes were dynamically synchronized to the twin system in real time, forming a continuously updated multi-source time-series data stream.

[0131] Based on the collected data, the optimized time-series graph neural network proposed in this paper was used to predict the evolution of soil water potential and crop water requirements for each plot over the next seven days. The network structure incorporates a spatiotemporal interaction weight modulation mechanism, a dynamic adjustment mechanism for external sensitivity, and a cycle-aware time encoding mechanism. The relevant structural parameters and hyperparameter configurations were dynamically optimized using the Red-billed Blue Magpie Optimization Algorithm. The optimized network outperformed traditional single-time series prediction models in terms of prediction accuracy and model stability, accurately capturing the changing trends in winter wheat water requirements during different growth stages.

[0132] Based on predicted future water demand trends and the base's real-time water supply capacity, including pump station outflow, channel capacity, and water source scheduling plans, irrigation control instructions are generated for each plot of land for the next week. Each instruction specifies the irrigation start time, duration, water volume, and sequence for each plot, ensuring optimal allocation and utilization of irrigation resources. The entire control instruction set is automatically synchronized to the farmland intelligent irrigation system terminal for scheduled execution. Simultaneously, feedback data from the irrigation process continues to flow back to the digital twin system in real time for subsequent predictions and strategy adjustments, forming a complete closed-loop intelligent management system.

[0133] In actual operation, the system of this invention experienced a typical drought from early to mid-March 2025. Traditional manual irrigation methods, due to delayed response, resulted in crop water stress rates as high as 22%, significantly reducing yields. However, the system of this invention, by predicting trends for the next seven days, rationally arranged irrigation plans in advance, effectively avoiding water stress and significantly improving wheat growth consistency and ultimate yield.

[0134] Specific data are as follows: While controlling water volume, the system reduces total water consumption by approximately 19.8% compared to traditional methods; the incidence of crop water stress has dropped from 22% to 5%; yields have increased by 8.4%; and the number of manual interventions has been reduced by 65%, significantly reducing labor costs. Furthermore, because the system dynamically adjusts irrigation strategies based on water supply, it effectively staggers irrigation during peak water pressure periods, avoiding pump station overloads and channel overflows, ensuring the stability and safety of the irrigation system.

[0135] Table 1 Comparison between the intelligent irrigation system of the present invention and traditional manual irrigation methods

[0136]

[0137]

[0138] As can be seen in Table 1, the intelligent farmland irrigation system proposed in this invention significantly outperforms traditional manual irrigation methods across all indicators. First, in terms of irrigation response time, traditional manual irrigation, due to its reliance on manual inspections and empirical judgment, has an average response time of 72 hours, significantly lagging behind actual changes in soil water potential and crop water requirements, easily leading to water stress or resource waste. However, the system proposed in this invention, through future trend prediction based on digital twins and time-series graph neural networks, can complete a response within 24 hours, a 66.7% reduction compared to traditional methods, significantly improving the real-time and accuracy of irrigation decisions.

[0139] In terms of total water consumption, traditional irrigation methods suffer from redundant irrigation and local over-irrigation in water supply arrangements, resulting in an overall water consumption of up to 10,200 tons. However, this invention accurately predicts the future water demand of each plot and generates optimized irrigation control instructions based on real-time water supply conditions. This ultimately reduces total water consumption to 8,180 tons, a reduction of approximately 19.8% compared to traditional methods. This result demonstrates that this invention not only improves irrigation water utilization efficiency but also effectively alleviates water resource shortages.

[0140] The incidence of crop water stress is a key indicator of irrigation strategy success. Due to delayed response and insufficient irrigation precision, traditional irrigation results in a 22% incidence of water stress in wheat populations, leading to uneven growth and reduced yields. However, with the application of this invention, the incidence of water stress dropped to 5%, a 77% reduction. This demonstrates the importance of precise prediction and intelligent regulation in ensuring a consistent crop growth environment, effectively improving the quality and uniformity of crop growth.

[0141] In terms of wheat yield, the average yield per mu under traditional irrigation is 455 kg / mu. However, through scientific irrigation scheduling and precise water supply, the present invention has increased wheat yield to 493 kg / mu, an increase of 8.4%. This yield increase directly verifies the effectiveness of the intelligent irrigation system in optimizing water management throughout the crop growth process, and also brings significant economic benefits to farm operators.

[0142] Traditional irrigation methods require frequent on-site inspections and manual activation and deactivation of water valves, averaging 20 interventions per week. This not only consumes significant human resources but also increases operational uncertainty. However, with the system of this invention, fully automated control instructions pushed and executed by intelligent terminals, only seven interventions are required per week, a 65% reduction. This significantly reduces labor input and operational costs, while also enhancing the intelligence and automation of farmland management.

[0143] In traditional systems, due to irrational irrigation planning, excessive water flow often leads to high pump station pressure and channel overflows. This phenomenon occurs approximately three times per week, wasting water resources and posing equipment damage and safety risks. However, the optimized control system of this invention, through staggered scheduling and rational water allocation, has reduced peak pump station overflows to zero, completely eliminating this safety hazard and ensuring the long-term stable operation of the irrigation system.

[0144] Based on the above analysis, the present invention is not only superior to traditional artificial irrigation methods in terms of irrigation response speed, water use efficiency, and crop growth guarantee, but also shows significant advantages in reducing human intervention and improving system operation safety and sustainability, fully verifying the great value and promotion potential of the present invention in actual farmland intelligent irrigation applications.

[0145] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A smart farmland irrigation decision-making method based on digital twins, characterized by: The steps include: S1. Collect multi-source data of farmland environment, pre-process the multi-source data, and generate a time series farmland environment dataset; S2. Based on the time-series farmland environmental dataset, a farmland digital twin model is constructed. The real-time collected environmental data is dynamically input into the farmland digital twin model to update the attribute information of nodes and edges. S3. Based on the updated farmland digital twin model, a time series graph is constructed, time series changes are encoded into a graph structure, and input into a time series graph neural network for modeling. The structural parameters and training hyperparameters of the time series graph neural network are preliminarily set. S4. Apply the red-billed blue magpie optimization algorithm to optimize the structural parameters and training hyperparameters of the time series graph neural network to obtain the optimal parameter configuration of the time series graph neural network; S5. Using the optimized time series graph neural network, the soil water potential evolution trend and crop water demand trend of each farmland plot in multiple future time steps are predicted, and the predicted water demand value of each plot corresponding to the time step is output; S6. Based on the prediction results and combined with the farmland water resources supply conditions, generate farmland irrigation control instructions, including the irrigation start time, irrigation duration, irrigation water volume and irrigation sequence for each plot; S7. During the irrigation process, the soil moisture changes and crop status changes of each plot are monitored in real time, feedback data is collected, and the feedback data is input into the farmland digital twin model. The attribute values ​​of the nodes and edges are corrected in real time, the twin status is updated, and the corrected farmland digital twin model is used for the prediction and optimization cycle of the next cycle.

2. The intelligent farmland irrigation decision-making method based on digital twin according to claim 1 is characterized in that: The multi-source data specifically includes soil moisture data, meteorological data, crop growth status data and plot geographic location information, which are used to build a digital twin model of farmland and support dynamic modeling and prediction of time series graph neural networks.

3. The intelligent farmland irrigation decision-making method based on digital twin according to claim 1 is characterized in that: The preprocessing of multi-source data specifically includes data cleaning, missing value filling, outlier removal and normalization processing, which is used to improve the quality of multi-source data.

4. The intelligent farmland irrigation decision-making method based on digital twin according to claim 1 is characterized in that: The S2 specifically includes: S21. Based on the time series farmland environment dataset, the farmland plot set is recorded as the node set V = {v1, v2, ..., v n }, the physical connection between farmland plots is recorded as edge set E = {e ij |v i ,v j ∈V}, construct the initial farmland graph G=(V,E), where v i represents the i-th farmland plot, i∈{1,2,…,n}, n is the total number of farmland plot nodes, e ij Represents node v i With node v j There is a direct physical connection between S22. Define each node v i The eigenvector x i =[s i ,t i ,c i ], where s i For plot v i Soil moisture value, t i For plot v i The current temperature value, c i For plot v i Corresponding crop growth status index value; S23. Define each edge e ij The eigenvalue w ij , representing the plot v i With plot v j The water flow trend between ij Obtained through comprehensive calculation based on the height difference, soil permeability and distance between plots; S24, the farmland environment data collected in real time is streamed Input into the initial farmland map G, where Represents plot v i At time t, the observation data, i∈{1,2,…,n}, updates the node v i The eigenvector of and edge ij The eigenvalue of S25, through the node feature vector and edge eigenvalues Continuous dynamic update to form a real-time evolving farmland digital twin model G t =(V,E,X t ,W t ),in S26, update the farmland digital twin model G t Serves as the basis for timing graph construction and timing graph neural network input.

5. The method for intelligent farmland irrigation decision-making based on digital twin according to claim 4, characterized in that: The S3 specifically includes: S31. In the updated farmland digital twin model G t Based on this, we define the node set V, edge set E, and node feature matrix X t , edge weight matrix W t , construct a time series diagram with time series T S32, at each time step t k On the top, extract node v i Features and edge weights Graph convolution operations are used to extract spatial features, and time recursive units are used to model temporal characteristics; S33, introduce the spatiotemporal dual attention mechanism, perform weighted fusion of spatial features and temporal features, and set the spatiotemporal interaction weight modulation coefficient λ st , the fusion feature is expressed as: in, Represents node v i At time step t k The fused spatiotemporal joint feature representation vector, For node v i At time step t k The extracted spatial feature representation, It is the time feature representation; S34, in the process of graph structure construction, based on the external factor dataset F t Dynamically adjust edge weights and define external sensitivity coefficient β ext , updated edge weights for: in, Indicates the rate of change of external factors on plot v i With v j The impact value of the relationship; S35. In the time coding process, a cycle-aware time coding method is used to encode each time step t. k Generate periodic encoding vector And introduce the period weight coefficient γ per , and obtain the improved temporal feature representation: in, For the tth k The time encoding vector after the fusion of time steps, is the absolute time code, Encode the cycle time; S36. Preliminary setting of structural parameters of the temporal graph neural network, including the number of graph convolution layers L g , time step expansion number m, node feature dimension d v , the number of hidden layer neurons h, and the newly added spatiotemporal interaction weight modulation coefficient λ st , external sensitivity coefficient β ext , period weight coefficient γ per , and set the training hyperparameters, including learning rate α, batch size b, and optimizer parameter set θ, to complete the initialization modeling of the timing graph neural network.

6. The intelligent farmland irrigation decision-making method based on digital twin according to claim 1 is characterized in that: The S4 specifically includes: S41, modulating the coefficient λ with the spatiotemporal interaction weight st , external sensitivity coefficient β ext , period weight coefficient γ per As the optimization object, construct the parameter set to be optimized Θ new ={λ st ,β ext ,γ per }; S42, using a population generation method based on a segmented initialization strategy to initialize the red-billed blue magpie optimization algorithm population, each red-billed blue magpie individual Θ new,p Randomly select initial values ​​from the divided reasonable parameter intervals, where p is the individual number of the red-billed blue magpie in the population; S43. Define comprehensive fitness function Using parameter Θ new,p The prediction error of the trained time series graph neural network on the validation set is weighted with the water resource utilization efficiency as the evaluation criteria: in, Configure Θ for parameters new,p The predicted output of the next i-th sample, y i is the true label of sample i, N is the number of samples in the validation set, η(Θ new,p ) is the water resource utilization efficiency corresponding to the farmland irrigation strategy derived from the prediction results, ω1 and ω2 are weighting coefficients; S44. According to the foraging behavior rules of the red-billed blue magpie, the individual Θ of the red-billed blue magpie population new,p To perform a location update: in, is the red-billed blue magpie individual with the best fitness in the tth generation, is another randomly selected red-billed blue magpie individual, r1 and r2 are random numbers in the range of [0,1]. is the updated parameter configuration vector of the red-billed blue magpie individual in the p-th population of the t+1 generation, is the updated parameter configuration vector of the red-billed blue magpie individual in the p-th population of the t-th generation; S45. Introducing a dynamic foraging radius adjustment mechanism to dynamically scale the search step size based on the fitness ranking of individual blue magpies in the population, expanding the exploration range in the early stages of the search and reducing the step size in the later convergence stages, thereby improving the global and local convergence of the search. S46. Introduce an individual variation mechanism based on multiple foraging strategies. When there is no fitness improvement for several consecutive generations, randomly adjust the single parameter λ of some low-fitness red-billed blue magpie individuals. st , β ext or γ per Perform disturbance mutation to prevent the population from falling into local optimality; S47, when the maximum number of iterations T is reached max Or after the fitness converges, the optimization process is terminated and the optimal parameter configuration is output The optimal parameter configuration is applied to the time series graph neural network for predictive modeling of soil water potential and crop water demand trends in farmland plots.

7. The method for intelligent farmland irrigation decision-making based on digital twin according to claim 1, characterized in that: The S5 specifically includes: S51. Based on the real-time updated farmland digital twin model, extract the node feature information of each plot at the current moment and the edge weight information between plots; S52, using the optimized temporal graph neural network, inputting node feature information and the relationship between plots, and performing feature encoding using the optimized structural parameters and the newly added fusion mechanism; S53, based on the input current time data, recursively predict the state changes of each farmland plot in multiple future time steps through the optimized time series graph neural network; S54. At each future time step, output the soil water potential evolution trend prediction results corresponding to each block, and record the dynamic changes of the soil water status of each block over time; S55, based on the predicted soil water potential trend, further derive and output crop water requirement trend data for each plot at each future time step; S56. Arrange the predicted water demand values ​​corresponding to each block at multiple future time steps to generate a farmland water demand prediction dataset.

8. The method for intelligent farmland irrigation decision-making based on digital twin according to claim 1, characterized in that: The S6 specifically includes: S61. Based on the future water demand prediction results output by the optimized time series graph neural network, extract the water demand trend data of each farmland plot at each future time step; S62. Collect data on water supply conditions in the farmland area in the current and foreseeable time period, including water supply volume, pump station operating status, and channel delivery capacity; S63. Based on water demand trend data and water resource supply conditions, conduct a water demand matching analysis at the plot level to determine a feasible irrigation start time period and preliminary water allocation plan; S64. Develop irrigation priorities for each plot based on the urgency of water demand and water supply capacity limitations, and generate an irrigation ranking list; S65. Determine the specific irrigation start time, irrigation duration, and irrigation water volume for each plot according to the irrigation ranking list, and form a specific plot irrigation control instruction set; S66. Integrate the irrigation control instructions of each plot to generate a global farmland irrigation control plan and execute the scheduling operation.

Images