Agricultural irrigation decision optimization method and system, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of agricultural irrigation, and provides an agricultural irrigation decision optimization method and system, electronic equipment and a storage medium, and the method comprises the steps: data pre-collection, simulation model construction, analog simulation, difference calculation, auxiliary model construction, decision model construction, decision generation, decision result prediction, decision optimization and decision implementation. According to the method, the simulation result is optimized by using the auxiliary model trained by the difference value between the AquaCrop model and the actual data, so that the simulation effect of the irrigation decision is improved, and the reliability of the decision is improved; through the constructed De-POMDP model, dynamic changes of environmental conditions and crop growth stages are flexibly responded, the processing efficiency of the model is improved through double-stage decision, and consumption of water resources and computing resources is reduced.

Description

Technical Field

[0001] This invention relates to the field of agricultural irrigation technology, and in particular to an agricultural irrigation decision optimization method, system, electronic device, and storage medium. Background Technology

[0002] Agricultural water use accounts for approximately 85% of available freshwater resources. However, climate change and global warming are severely impacting rainfall patterns and water availability. Different crops have varying water requirements, necessitating optimized irrigation decisions to ensure crop yield and quality under these complex conditions. Traditional irrigation methods often rely on fixed time intervals or empirical plans, failing to adequately consider dynamic environmental changes or the real-time needs of crop growth. This results in low water use efficiency and severely limited system flexibility and adaptability. In recent years, reinforcement learning has been increasingly introduced into agricultural management, achieving significant progress, particularly in irrigation and fertilization management.

[0003] Despite significant progress in these studies, they still have obvious limitations in solving real-world agricultural irrigation problems. First, most studies employ fixed decision intervals, making it difficult for the system to flexibly respond to dynamic changes in environmental conditions and crop growth stages. Static decision-making strategies based on fixed time intervals may miss optimal irrigation times or make unnecessary and frequent decisions, leading to water waste and excessive consumption of computational resources. Second, current research primarily uses single-agent models, which are insufficient to handle multi-objective decisions regarding irrigation time and volume. These decision objectives are often interdependent and dynamically changing, making it difficult for single-agent systems to effectively handle these complex tasks, resulting in inefficient decision-making. Finally, the high-dimensional state space in complex environments, such as weather conditions, soil moisture, and crop growth stages, significantly increases the training time and computational resources required for the models, slowing down the convergence speed of the strategies and further limiting the practical application of these models. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an agricultural irrigation decision optimization method, system, electronic device, and storage medium, which solves the problems of large errors between existing simulation model results and real data, as well as the large consumption of water and computing resources in decision generation based on fixed decision space and single agent models.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] An agricultural irrigation decision optimization method includes:

[0007] Data was collected from the target irrigation area to obtain soil data, climate data, crop information, and crop data.

[0008] The soil data, climate data, and crop information are input into the AquaCrop model to obtain a crop simulation model;

[0009] The crop simulation model is used to simulate the crop yield and irrigation cost of the target irrigation area to obtain model prediction data;

[0010] Calculate the difference between the model's predicted data and the crop data to obtain an error dataset;

[0011] The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model.

[0012] Construct a Dec-POMDP model; the Dec-POMDP model includes: an irrigation time decision channel and an irrigation quantity decision channel;

[0013] The soil data, climate data, and crop information are input into the Dec-POMDP model to generate irrigation decisions, resulting in irrigation time and irrigation amount decisions.

[0014] The irrigation time decision and the irrigation amount decision are input into the optimized prediction model for prediction, and the predicted yield and predicted cost are obtained.

[0015] Based on the predicted yield, the predicted cost, the crop yield, and the irrigation cost, the Dec-POMDP model is iteratively optimized using the IPPO algorithm to obtain irrigation time optimization decisions and irrigation amount optimization decisions;

[0016] Irrigation treatment is carried out on the target irrigation area based on the irrigation time optimization decision and the irrigation volume optimization decision.

[0017] Preferably, data is collected from the target irrigation area to obtain soil data, climate data, crop information, and crop data, including:

[0018] The soil data are obtained by collecting the total available soil water, soil water deficit, relative water deficit in the root zone, and basic soil texture parameters of the target irrigation area. The basic soil texture parameters include: soil type and porosity.

[0019] The climate data is obtained by collecting rainfall data, crop evapotranspiration data, temperature, and relevant meteorological parameters in the target irrigation area; the relevant meteorological parameters include: relative humidity, wind speed, solar radiation, and CO2 concentration.

[0020] The crop type, crop variety, and planting parameters of the target irrigation area are collected to obtain the crop information;

[0021] The crop data are obtained by statistically analyzing the canopy coverage, aboveground biomass, growth stage, total water consumption, and seasonal water surplus of crops within the target irrigation area.

[0022] Preferably, data collection is performed on the target irrigation area to obtain soil data, climate data, crop information, and crop data, and the method further includes:

[0023] Outlier removal and missing value filling are performed sequentially on the soil data, the climate data, the crop information, and the crop data;

[0024] Convert the soil type, crop type, and crop variety into unique thermal codes;

[0025] The porosity, air temperature, and aboveground biomass were pretreated using the min-max normalization method.

[0026] The Z-score normalization method was used to analyze the total available soil water, the soil water deficit, the rainfall data, the crop evapotranspiration data, and the... The concentration, wind speed, total water consumption, and seasonal remaining water volume are pretreated.

[0027] The solar radiation and canopy coverage were preprocessed using the max normalization method.

[0028] The relative water deficit in the root zone and the relative humidity are mapped proportionally to the unit interval.

[0029] Preferably, the improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including:

[0030] The improved GRU network is constructed as follows: the improved GRU network includes an input layer, an encoding layer, and an output layer connected in sequence; the input layer includes an embedding layer, a first fully connected layer, a linear projection layer, and a feature vector fusion layer; the encoding layer includes a first improved GRU with embedded feature adaptive gating, a second improved GRU with bidirectional stacking, a temporal attention layer, a feature attention layer, a residual connection layer, a LayerNorm normalization layer, a Dropout layer, and a second fully connected layer connected in sequence; the embedding layer is connected to the first fully connected layer; the first fully connected layer and the linear projection layer are respectively connected to the feature vector fusion layer;

[0031] The processed soil data, climate data, and crop information are categorized into static data and dynamic data.

[0032] The static data is transformed using the embedding layer and the first fully connected layer to obtain a static vector.

[0033] Preferably, the improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including:

[0034] The dynamic data is linearly mapped in time using the linear projection layer to obtain a dynamic vector, and the static vector and the dynamic vector are fused in time steps using the feature vector fusion layer to obtain an input sequence;

[0035] The first improved GRU is used to extract features from the input sequence to obtain the hidden state;

[0036] The second improved GRU is used to extract forward and backward dependencies from the hidden state to obtain bidirectional features;

[0037] The temporal attention layer is used to calculate the time step weights and normalize the bidirectional features to obtain global temporal features;

[0038] The feature attention layer is used to calculate the feature contribution weights of the global temporal features to obtain the key feature vector.

[0039] The residual connection layer is used to add the key feature vector and the residual output of the second improved GRU to obtain the gradient reduction vector;

[0040] The gradient reduction vector is normalized and randomly discarded using the LayerNorm normalization layer and Dropout layer, respectively, to obtain regularized features;

[0041] The output difference branch of the output layer is used to perform dimensionality reduction mapping on the regularized features to obtain the predicted output difference;

[0042] The regularized features are dimensionality-reduced and mapped using the cost difference branch of the output layer to obtain the predicted cost difference.

[0043] Preferably, the improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including:

[0044] Construct the production error loss; the expression for the production error loss is: ;in, This refers to the production error loss; This represents the total time step; Weights for time-series stages; Huber loss function; For model predictions Time step output difference; for The difference between the actual output at each time step;

[0045] Construct cost error loss; the expression for the cost error loss is: ;in, This refers to the cost error loss; Cost weighting; for The soil moisture deficit at the time step; For model predictions Time step cost difference; for The actual cost difference of each time step;

[0046] Construct a regular expression; the expression for the regular expression is: ;in, For the regularization term; The number of samples; , The first Predicted values ​​for production difference and cost difference for each sample; , These are the mean error of historical output difference and the mean error of historical cost difference, respectively.

[0047] Integrating the production error loss, the cost error loss, and the regularization term, we obtain the overall loss function; the expression for the overall loss function is: ;in, The overall loss function; This is the dynamic balance coefficient; The regularization coefficient is used.

[0048] The improved GRU network is iterated using the overall loss function based on the predicted cost difference and the error dataset to obtain the optimized prediction model.

[0049] Preferably, constructing the Dec-POMDP model includes:

[0050] Construct a state space; the state space includes: , Month, DAP, IrrCum, IrrSur, CC, B, GrowthStage, Dep, TAW, Depletion, Rain1, Rain2, Rain3, ET1, ET2, ET3; among them, Month, DAP, IrrCum, IrrSur, CC, B, GrowthStage, Dep, TAW, Depletion, Rain1, Rain2, Rain3, ET1, ET2, and ET3 represent the date, month, total number of days since planting, total water consumption, seasonal remaining water, canopy coverage, aboveground biomass, growth stage, relative water deficit in the root zone, total available soil water, soil water deficit, average daily rainfall over the past 7 days, total rainfall, yesterday's rainfall, average daily reference crop evapotranspiration over the past 7 days, total crop evapotranspiration, and yesterday's reference crop evapotranspiration.

[0051] Determine the number of agents; the number of agents is 2.

[0052] Construct an action space; the expression of the action space includes: and ;in, To select any day from day 0 to day 13 after the current decision as the irrigation time; For the selected irrigation amount;

[0053] Construct an observation space; the expression for the observation space is: and ;in, This refers to the observation space of the irrigation time decision channel; This refers to the observation space of the irrigation volume decision channel;

[0054] Define the state transition probability;

[0055] Construct a reward function; the expression of the reward function is: ;in, The reward function; For time step The state of being; For time step Joint actions at the location; For crop prices; The unit cost of irrigation; The total yield of the crop on the harvest day; For the first Daily irrigation volume; The fixed non-irrigated production cost for this round at the end of the simulation;

[0056] Define a discount factor; the expression for the discount factor is: ;in, Let the objective function be the policy objective function; For corresponding actions and state Expectations; The discount factor; For instant rewards;

[0057] The Dec-POMDP model is obtained by integrating the state space, the number of agents, the action space, the observation space, the state transition probability, the reward function, and the discount factor.

[0058] Preferably, an agricultural irrigation decision optimization system includes:

[0059] The data pre-acquisition module is used to collect data from the target irrigation area to obtain soil data, climate data, crop information, and crop data.

[0060] The simulation model building module is used to input the soil data, climate data, and crop information into the AquaCrop model to obtain a crop simulation model.

[0061] The simulation module is used to simulate the crop yield and irrigation cost of the target irrigation area using the crop simulation model, and obtain model prediction data.

[0062] The difference calculation module is used to calculate the difference between the model prediction data and the crop data to obtain an error dataset.

[0063] An auxiliary model building module is used to train the improved GRU network using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model.

[0064] The decision model construction module is used to construct the Dec-POMDP model; the Dec-POMDP model includes: an irrigation time decision channel and an irrigation amount decision channel;

[0065] The decision generation module is used to input the soil data, climate data, and crop information into the Dec-POMDP model to generate irrigation decisions, resulting in irrigation time and irrigation amount decisions.

[0066] The decision result prediction module is used to input the irrigation time decision and the irrigation amount decision into the optimized prediction model to make predictions and obtain predicted yield and predicted cost.

[0067] The decision optimization module is used to iteratively optimize the Dec-POMDP model using the IPPO algorithm based on the predicted yield, the predicted cost, the crop yield, and the irrigation cost, to obtain the irrigation time optimization decision and the irrigation amount optimization decision.

[0068] The decision implementation module is used to perform irrigation treatment on the target irrigation area based on the irrigation time optimization decision and the irrigation amount optimization decision.

[0069] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to execute the aforementioned method for formulating qualification standards for operation ticket personnel based on multivariate knowledge analysis.

[0070] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned method for formulating qualification standards for operation ticket personnel based on multivariate knowledge analysis.

[0071] The present invention discloses the following technical effects:

[0072] This invention provides an agricultural irrigation decision optimization method, system, electronic device, and storage medium. By using an auxiliary model trained with the difference between the AquaCrop model and actual data to optimize simulation results, it solves the problem of large errors between existing simulation model results and real data, and achieves optimization of irrigation decision simulation results. Through the constructed Dec-POMDP model, it solves the problem of large water and computing resource consumption in existing decision-making models based on fixed decision space and single agent model, and realizes dynamic response and two-stage processing functions for environmental conditions and crop growth stages. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 This is a schematic diagram of the agricultural irrigation decision optimization process provided in an embodiment of the present invention;

[0075] Figure 2 This is a schematic diagram of agricultural irrigation decision optimization provided in an embodiment of the present invention;

[0076] Figure 3This is a schematic diagram of the feature extraction and construction process provided in an embodiment of the present invention;

[0077] Figure 4 This is a schematic diagram of the feature processing flow provided in an embodiment of the present invention;

[0078] Figure 5 This is a schematic diagram of the model iteration process provided in an embodiment of the present invention. Detailed Implementation

[0079] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] The purpose of this invention is to provide an agricultural irrigation decision optimization method, system, electronic device, and storage medium to solve the problems of large errors between existing simulation model results and real data, as well as the high consumption of water and computing resources in decision generation based on fixed decision space and single agent models.

[0081] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] Figure 1 This is a schematic diagram of the agricultural irrigation decision optimization process provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of agricultural irrigation decision optimization provided in an embodiment of the present invention, such as... Figure 1 and Figure 2 As shown, the present invention provides an agricultural irrigation decision optimization method, comprising:

[0083] Step 100: Collect data from the target irrigation area to obtain soil data, climate data, crop information, and crop data;

[0084] Step 200: Input the soil data, climate data, and crop information into the AquaCrop model to obtain a crop simulation model;

[0085] Step 300: Use the crop simulation model to simulate the crop yield and irrigation cost of the target irrigation area to obtain model prediction data;

[0086] Step 400: Calculate the difference between the model prediction data and the crop data to obtain the error dataset;

[0087] Step 500: Train the improved GRU network using the error dataset to obtain the trained auxiliary calibration model. Integrate the crop simulation model and the auxiliary calibration model to obtain the optimized prediction model.

[0088] Step 600: Construct the Dec-POMDP model; the Dec-POMDP model includes: irrigation time decision channel and irrigation amount decision channel;

[0089] Step 700: Input the soil data, the climate data, and the crop information into the Dec-POMDP model to generate irrigation decisions, and obtain irrigation time decisions and irrigation amount decisions;

[0090] Step 800: Input the irrigation time decision and the irrigation amount decision into the optimized prediction model to make a prediction, and obtain the predicted yield and predicted cost;

[0091] Step 900: Based on the predicted yield, the predicted cost, the crop yield, and the irrigation cost, the IPPO algorithm is used to iteratively optimize the Dec-POMDP model to obtain the irrigation time optimization decision and the irrigation amount optimization decision;

[0092] Step 1000: Irrigate the target irrigation area according to the irrigation time optimization decision and the irrigation amount optimization decision.

[0093] Specifically, data is collected from the target irrigation area to obtain soil data, climate data, crop information, and crop data, including:

[0094] The soil data are obtained by collecting the total available soil water, soil water deficit, relative water deficit in the root zone, and basic soil texture parameters of the target irrigation area. The basic soil texture parameters include: soil type and porosity.

[0095] The climate data is obtained by collecting rainfall data, crop evapotranspiration data, temperature, and relevant meteorological parameters in the target irrigation area; the relevant meteorological parameters include: relative humidity, wind speed, solar radiation, and CO2 concentration.

[0096] The crop type, crop variety, and planting parameters of the target irrigation area are collected to obtain the crop information;

[0097] The crop data are obtained by statistically analyzing the canopy coverage, aboveground biomass, growth stage, total water consumption, and seasonal water surplus of crops within the target irrigation area.

[0098] Furthermore, data is collected from the target irrigation area to obtain soil data, climate data, crop information, and crop data, including:

[0099] Outlier removal and missing value filling are performed sequentially on the soil data, the climate data, the crop information, and the crop data;

[0100] Convert the soil type, crop type, and crop variety into unique thermal codes;

[0101] The porosity, air temperature, and aboveground biomass were pretreated using the min-max normalization method.

[0102] The Z-score normalization method was used to analyze the total available soil water, the soil water deficit, the rainfall data, the crop evapotranspiration data, and the... The concentration, wind speed, total water consumption, and seasonal remaining water volume are pretreated.

[0103] The solar radiation and canopy coverage were preprocessed using the max normalization method.

[0104] The relative water deficit in the root zone and the relative humidity are mapped proportionally to the unit interval.

[0105] refer to Figure 3 The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including:

[0106] Step 501: Construct the improved GRU network; the improved GRU network includes: an input layer, an encoding layer, and an output layer connected in sequence; the input layer includes: an embedding layer, a first fully connected layer, a linear projection layer, and a feature vector fusion layer; the encoding layer includes: a first improved GRU with embedded feature adaptive gating, a second improved GRU with bidirectional stacking, a temporal attention layer, a feature attention layer, a residual connection layer, a LayerNorm normalization layer, a Dropout layer, and a second fully connected layer connected in sequence; the embedding layer is connected to the first fully connected layer; the first fully connected layer and the linear projection layer are respectively connected to the feature vector fusion layer;

[0107] Step 502: Divide the processed soil data, climate data, and crop information into static data and dynamic data;

[0108] Step 503: Use the embedding layer and the first fully connected layer to perform dimensionality transformation on the static data to obtain a static vector.

[0109] refer to Figure 4The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including:

[0110] Step 504: The dynamic data is linearly mapped according to time sequence using the linear projection layer to obtain a dynamic vector, and the static vector and the dynamic vector are fused according to time step using the feature vector fusion layer to obtain an input sequence;

[0111] Step 505: Use the first improved GRU to extract features from the input sequence to obtain the hidden state;

[0112] Step 506: Use the second improved GRU to perform forward dependency extraction and backward dependency extraction on the hidden state to obtain bidirectional features;

[0113] Step 507: Calculate and normalize the time step weights of the bidirectional features using the temporal attention layer to obtain global temporal features;

[0114] Step 508: Calculate the feature contribution weights of the global temporal features using the feature attention layer to obtain the key feature vector;

[0115] Step 509: Add the key feature vector and the residual output of the second improved GRU using the residual connection layer to obtain the gradient reduction vector;

[0116] Step 510: Normalize and randomly discard the gradient reduction vector using the LayerNorm normalization layer and Dropout layer respectively to obtain regularized features;

[0117] Step 511: Use the output difference branch of the output layer to perform dimensionality reduction mapping on the regularized features to obtain the predicted output difference;

[0118] Step 512: Use the cost difference branch of the output layer to perform dimensionality reduction mapping on the regularized features to obtain the predicted cost difference.

[0119] refer to Figure 5 The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including:

[0120] Step 513: Construct the production error loss; the expression for the production error loss is: ;in, This refers to the production error loss; This represents the total time step; Weights for time-series stages; Huber loss function; For model predictions Time step output difference; for The difference between the actual output at each time step;

[0121] Step 514: Construct the cost error loss; the expression for the cost error loss is: ;in, This refers to the cost error loss; Cost weighting; for The soil moisture deficit at the time step; For model predictions Time step cost difference; for The actual cost difference of each time step;

[0122] Step 515: Construct a regular expression; the expression for the regular expression is: ;in, For the regularization term; The number of samples; , The first Predicted values ​​for production difference and cost difference for each sample; , These are the mean error of historical output difference and the mean error of historical cost difference, respectively.

[0123] Step 516: Integrate the production error loss, the cost error loss, and the regularization term to obtain the overall loss function; the expression of the overall loss function is: ;in, The overall loss function; This is the dynamic balance coefficient; The regularization coefficient is used.

[0124] Step 517: Based on the predicted cost difference and the error dataset, iterate the improved GRU network using the overall loss function to obtain the optimized prediction model.

[0125] Furthermore, the Dec-POMDP model is constructed, including:

[0126] Construct a state space; the state space includes: , Month, DAP, IrrCum, IrrSur, CC, B, GrowthStage, Dep, TAW, Depletion, Rain1, Rain2, Rain3, ET1, ET2, ET3; among them, Month, DAP, IrrCum, IrrSur, CC, B, GrowthStage, Dep, TAW, Depletion, Rain1, Rain2, Rain3, ET1, ET2, and ET3 represent the date, month, total number of days since planting, total water consumption, seasonal remaining water, canopy coverage, aboveground biomass, growth stage, relative water deficit in the root zone, total available soil water, soil water deficit, average daily rainfall over the past 7 days, total rainfall, yesterday's rainfall, average daily reference crop evapotranspiration over the past 7 days, total crop evapotranspiration, and yesterday's reference crop evapotranspiration.

[0127] Determine the number of agents; the number of agents is 2.

[0128] Construct an action space; the expression of the action space includes: and ;in, To select any day from day 0 to day 13 after the current decision as the irrigation time; For the selected irrigation amount;

[0129] Construct an observation space; the expression for the observation space is: and ;in, This refers to the observation space of the irrigation time decision channel; This refers to the observation space of the irrigation volume decision channel;

[0130] Define the state transition probability;

[0131] Construct a reward function; the expression of the reward function is: ;in, The reward function; For time step The state of being; For time step Joint actions at the location; For crop prices; The unit cost of irrigation; The total yield of the crop on the harvest day; For the first Daily irrigation volume; The fixed non-irrigated production cost for this round at the end of the simulation;

[0132] Define a discount factor; the expression for the discount factor is: ;in, Let the objective function be the policy objective function; For corresponding actions and state Expectations; The discount factor; For instant rewards;

[0133] The Dec-POMDP model is obtained by integrating the state space, the number of agents, the action space, the observation space, the state transition probability, the reward function, and the discount factor.

[0134] Specifically, soil data for the irrigated area were collected, including total available water, water deficit, relative water deficit in the root zone, soil type, porosity, bulk density, pH, and organic matter. Relative water deficit in the root zone reflects the degree of soil water stress and is the ratio of actual water deficit in the root zone to total available water. Climate data were also collected, including rainfall (daily, hourly, and intensity), crop evapotranspiration (reference and actual), temperature, humidity, wind speed, solar radiation, and CO2 concentration. Crop evapotranspiration was calculated using the Penman-Monteith formula. Crop information included type, variety, sowing date, planting density, row and plant spacing, growth stage, and irrigation method. Crop data included canopy cover (measured by remote sensing or canopy analyzer), aboveground biomass (destructive sampling or non-destructive monitoring), growth stage, total water consumption, seasonal water surplus, leaf area index, and crop height.

[0135] Furthermore, AquaCrop, based on the dynamic response of crops to water, combines meteorological conditions, soil moisture movement, crop physiological characteristics, and management measures (such as irrigation and fertilization) to simulate the water balance and crop growth status of the crop-soil-atmosphere system. It has been widely applied in agricultural water resource management research across different climate zones and various crop types worldwide. To effectively integrate the AquaCrop model with the reinforcement learning framework, this embodiment utilizes the Python version of AquaCrop-OSPy and the open-source reinforcement learning environment building tool aquacrop-gym to develop a dual-channel heterogeneous agent environment, CropMultiAgentEnv, with a standard interface. This environment retains AquaCrop's physiological simulation capabilities while encapsulating the crop growth process using an OpenAIGym-style interface, enabling the reinforcement learning agent to continuously interact with the environment throughout the complete crop growth cycle, thereby dynamically optimizing irrigation strategies. During the reinforcement learning training phase, the system randomly selects daily meteorological data (such as maximum and minimum temperatures, precipitation, and solar radiation) from the dataset for one year and inputs it along with soil, crop, and management information into AquaCrop. During the simulation, each time step determines irrigation demand and growth changes based on meteorological conditions, soil moisture balance, and crop growth stage. The reinforcement learning agent makes real-time decisions on whether to irrigate and the amount of irrigation by comprehensively evaluating these factors, thereby optimizing water management under various climatic and soil conditions.

[0136] Preferably, soil data, climate data, and crop information are input into the AquaCrop model to obtain a crop simulation model. This model is used to simulate crop yield and irrigation cost in the target irrigation area to obtain predicted data. Then, the difference between the predicted data and the actual crop data is calculated to obtain an error dataset. First, outlier handling is performed on soil, climate, crop information, and crop data, and missing values ​​are filled in. Next, soil type, crop type, and crop variety are converted into unique thermal codes. Then, min-max normalization is used to preprocess porosity, air temperature, and aboveground biomass. Z-score normalization is then used to preprocess total available soil water, soil water deficit, rainfall, crop evapotranspiration, concentration, wind speed, total water consumption, and seasonal surplus water. After that, max normalization is used to preprocess solar radiation and canopy cover. Finally, the relative water deficit, relative humidity, and canopy cover in the root zone are mapped proportionally to unit intervals.

[0137] Furthermore, an improved GRU network is constructed, comprising an input layer, an encoding layer, and an output layer connected in sequence. The input layer includes an embedding layer, a first fully connected layer, a linear projection layer, and a feature vector fusion layer. The embedding layer is connected to the first fully connected layer; both the first fully connected layer and the linear projection layer are connected to the feature vector fusion layer. The encoding layer consists of the following components in sequence: a first improved GRU with embedded feature adaptive gating, a second improved GRU with bidirectional stacking, a temporal attention layer, a feature attention layer, a residual connection layer, a LayerNorm normalization layer, a Dropout layer, and a second fully connected layer. The preprocessed soil data, climate data, and crop information are divided into static and dynamic data. Static data includes soil, crop type, crop variety, and soil porosity; dynamic data includes total available soil water, soil water deficit, relative water deficit in the root zone, rainfall, crop evapotranspiration, temperature, humidity, wind speed, solar radiation, CO2 concentration, total water consumption, seasonal water surplus, canopy cover, aboveground biomass, and growth stage. The embedding layer processes categorical features in static data, transforming high-dimensional discrete features into low-dimensional dense vectors. The embedding dimension is automatically adjusted based on the number of features. First, the initial fully connected layer integrates categorical features and static numerical features, using the ReLU activation function to unify the dimension and generate a static base vector. Next, this vector is batch-normalized to reduce differences in feature dimensions, making model training more stable. Dynamic data is aligned with the time series and mapped dimension through a linear projection layer. Specifically, dynamic features are arranged into a daily time series, and a linear transformation matrix is ​​used to project the dynamic feature vector for each day. Finally, the static and dynamic vectors are summed one by one, and then fused using attention-weighted fusion to obtain the final input sequence.

[0138] Preferably, a production error loss function is constructed:

[0139]

[0140] Construction cost error loss:

[0141]

[0142] Constructing regular expressions:

[0143]

[0144] The overall loss function is obtained by integration:

[0145]

[0146] Furthermore, dynamic time-series data processing and vector fusion are performed: preprocessed dynamic data (such as soil moisture and rainfall) is transformed into fixed-dimensional dynamic vectors in chronological order through linear projection. Static data (such as soil type and crop variety) is also transformed into static vectors. The feature vector fusion layer integrates the static and dynamic vectors by time points and adjusts their weights according to the crop growth stage, thereby forming a sequence containing spatiotemporal information.

[0147] Improved GRU Feature Extraction: The improved GRU module receives the input sequence and uses feature adaptive gating to identify the impact of different factors (such as soil moisture and temperature) on the error, retaining features highly correlated with crop yield and cost, filtering noise, and then outputting the hidden state features. Next, a second improved GRU, bidirectionally stacked, performs bidirectional feature extraction on the hidden state: the forward GRU captures historical dependencies, and the backward GRU captures future associations, ultimately outputting a comprehensive feature that fuses bidirectional temporal information.

[0148] Attention Mechanism and Feature Enhancement: The temporal attention layer calculates the weights of bidirectional features at different time steps, focusing more on key stages of crop growth to enhance features at important time points and obtain global temporal features. Then, the feature attention layer evaluates the importance of each factor in these global temporal features and selects the key vectors that have the greatest impact on prediction error.

[0149] Feature optimization and regularization: The residual connection layer adds the key feature vector to the original features output by the second improved GRU to alleviate gradient vanishing and obtain a more stable gradient reduction vector. The LayerNorm normalization layer normalizes the gradient reduction vector across samples to unify the feature distribution; the Dropout layer randomly discards some neurons to avoid overfitting. After compressing the feature dimension, the second fully connected layer yields standardized regularized features.

[0150] Difference Prediction and Model Integration: The output layer contains two branches: a yield difference branch predicts the difference between the crop simulation model's predicted yield and the actual yield; and a cost difference branch predicts the difference in irrigation costs. During training, using the error dataset as the target, backpropagation is used to optimize the network parameters, enabling the auxiliary calibration model to accurately capture the systematic errors of the simulation model. Finally, the output of the crop simulation model is superimposed with the difference predicted by the auxiliary calibration model for calibration, forming an optimized prediction model.

[0151] Preferably, the irrigation task is defined as follows:

[0152] 1) Abstraction of the irrigation decision problem:

[0153] This embodiment abstracts the irrigation decision problem into a problem with a decision-time dimension. Decision-making channel dimension and feature dimensions The three-dimensional decision space, composed of these dimensions, is described below: Decision Time Dimension This indicates the time point in time when the decision occurred. Assuming the smallest unit of decision-making is the day, then... = , representing the An irrigation decision is made daily, which includes determining the amount of irrigation to be done and when to make the next irrigation decision.

[0154] Decision-making channel dimension This represents the different decision-making channels involved in the decision-making task. This embodiment sets... =2, including two heterogeneous decision-making channels: the first decision-making channel Focus on determining when to make the next irrigation decision; second decision channel The focus is then on determining the specific water usage for this irrigation. Each channel makes its own independent decision based on its own set of relevant observation features.

[0155] Feature Dimension The specific set of features used for decision-making mainly includes two categories: environmental condition features and crop growth status features. Environmental condition features include meteorological information such as temperature, rainfall, and reference crop evapotranspiration, as well as soil moisture information such as total available soil water and soil water deficit. Crop growth status features include information related to crop physiological status, such as aboveground biomass, canopy cover, growth stage, and root zone water consumption.

[0156] 2) Definition of state information at any decision moment: at any decision moment Each decision channel acquires and utilizes the following state information: environmental condition information set. This includes various historical meteorological data, such as minimum and maximum temperatures, rainfall, and reference crop evapotranspiration, all of which directly affect crop growth and water requirements. In addition, soil moisture status is also a crucial factor influencing crop growth, including total available soil water and soil water deficit. Total available soil water represents the maximum amount of water the root zone soil can store and provide for crop absorption, while soil water deficit indicates the current degree of soil water deficiency. The ratio of the latter to the former is called the relative root zone water deficit, which measures the current extent of soil water shortage in the root zone.

[0157] Crop status information set This reflects the current growth status and actual water requirements of crops, specifically including key characteristics such as aboveground biomass, canopy cover, and crop growth stage. Aboveground biomass measures the growth of the aboveground parts of the crop; canopy cover indicates the degree to which the crop canopy covers the soil surface, directly affecting transpiration and water loss rates; and crop growth stage describes the current stage of the crop's growth cycle, such as seedling, flowering, or maturity.

[0158] 3) Formal definition of irrigation decision-making task: In the first... Decision-making moment The irrigation decision-making task defined in this embodiment is formally represented as follows:

[0159]

[0160] In the formula, Indicates the first At the decision-making moment, by the first The decision-making tasks performed by each decision-making channel; and These represent the environmental conditions and crop status information at the current moment, respectively. This indicates the specific feature information upon which the decision is based, obtained through the aforementioned feature set.

[0161] Specifically, the irrigation decision-making problem is modeled as a Dec-POMDP problem. This embodiment models the irrigation decision-making problem as a fully cooperative, decentralized, partially observable Markov Decision Process (Dec-POMDP). Dec-POMDP is a mathematical framework used to study the problem of choosing the order of actions under partial observability and environmental stochasticity. This framework allows multiple agents to make collaborative decisions in a shared environment, where each agent can only partially observe global state information. The Dec-POMDP model is... definition:

[0162] (State space): Describes all possible states of the simulated environment, including various physiological states of crop growth, soil moisture conditions, and weather conditions. Table 1 lists the specific state variables.

[0163] Table 1

[0164] state space definition day day Month moon DAP Total number of days since planting IrrCum Current total water consumption (mm / ha) IrrSur Seasonal remaining water volume (mm / ha) CC Canopy coverage (%) B Aboveground biomass (kg / ha) GrowthStage growth stage Dep Relative water deficit in the root zone (%) TAW Total available soil water (mm / ha) Depletion Soil moisture deficit (mm / ha) Rain1 Average daily rainfall over the past 7 days (mm / ha) Rain2 Total rainfall since planting (mm / ha) Rain3 Yesterday's rainfall (mm / ha) ET1 Average daily reference crop evapotranspiration over the past 7 days (mm / ha) ET2 Total evapotranspiration of reference crops since planting (mm / ha) ET3 Yesterday's reference crop evapotranspiration (mm / ha)

[0165] (Number of agents): express A collection of agents. In this embodiment, two heterogeneous agents are designed to undertake different tasks, each processed by a separate channel. The agent in the first channel is responsible for determining the irrigation time based on historical precipitation and reference crop evapotranspiration; the agent in the second channel determines the irrigation amount based on real-time information such as soil moisture and crop growth status.

[0166] (Action Space): Defines the set of operations that each agent can perform in the irrigation decision. In this decision problem, it is designed with two main decision dimensions: choosing the irrigation time from day 0 to day 13 after this decision; and determining the specific irrigation amount within the selected time, where the discrete depth irrigation amount is determined from the set... Choose from:

[0167]

[0168]

[0169] (Observation space): Each intelligent agent , In the global state Local observations below In this decision-making problem, the local observation space design of the two heterogeneous agents embodies a dual-channel characteristic, thus better handling the complexity of irrigation decisions. For the irrigation timing decision-making agent, observations include the historical weekly average daily precipitation, reference crop evapotranspiration, crop growth stage, root zone water consumption, and the total number of days since planting. These observations help the agent focus on long-term climatic and crop growth cycle factors influencing irrigation timing. On the other hand, the irrigation amount decision-making agent's observations cover current root zone water consumption, canopy cover, aboveground biomass, cumulative irrigation amount, remaining water supply, and crop growth stage. This allows the agent to adjust irrigation amount based on real-time information on soil moisture and the current crop status.

[0170]

[0171]

[0172] (State transition probability): Defines the probability of a given joint action of all agents. In the case of the current state Transition to the next state probability This includes responses to crop growth models and the effects of irrigation.

[0173] (Reward Function): In the multi-agent irrigation decision-making framework of this study, all agents share a single reward function. This function is used to evaluate the agent at each time step. The reward function is designed to balance crop yield and irrigation costs, thereby optimizing the overall system performance. At time step... Shared reward function Defined as:

[0174]

[0175] in, The reward function; For time steps The state of being; For time steps Joint actions at the location; For crop prices; The unit cost of irrigation; The total yield of the crop on the harvest day; For the first Daily irrigation volume; This represents the fixed non-irrigated production cost at the end of the simulation.

[0176] (Discount Factor): Used to calculate the current value of future rewards, reflecting the time-discounting effect on rewards. The agent uses parameters... Defined strategy From a local perspective Generate Actions And jointly optimize the discount accumulation reward:

[0177]

[0178] in, Let the objective function be the policy objective function; For corresponding actions and state Expectations; Let t be the discount factor for day t; For corresponding actions and state Instant rewards.

[0179] Preferably, this embodiment designs two heterogeneous intelligent agent channels with complementary functions:

[0180] Irrigation Timing Decision Channel: This agent channel acquires weather and soil information from the environment, including precipitation, reference crop evapotranspiration, total available soil water, soil moisture deficit, and time information such as crop growth stage and total number of days since planting, to determine the optimal irrigation decision time. Through reinforcement learning algorithms, the agent can mine potential decision patterns from the observation data and flexibly adjust irrigation strategies to adapt to different environmental conditions.

[0181] Irrigation Quantity Decision Channel: This agent channel focuses on analyzing real-time crop growth status, comprehensively considering factors such as crop growth stage, current soil moisture condition, and remaining water supply to dynamically adjust irrigation quantity. This channel utilizes a real-time data-driven strategy to effectively avoid over-irrigation or under-irrigation. The two agent channels collaborate by sharing key data, ensuring consistency and coordination of irrigation strategies. Building upon this, a globally shared reward function is introduced. This function, within a reinforcement learning framework, comprehensively considers multi-objective optimization of crop yield and water consumption, encouraging efficient cooperation and information sharing among agent channels, thereby improving overall system performance.

[0182] Optionally, Independent Proximal Policy Optimization (IPPO) is employed. The IPPO algorithm is an improved policy optimization method with significant advantages in handling heterogeneous agent problems. Its core idea is to independently optimize the policy of each agent, thereby effectively addressing the complex interactions and coordination problems between agents. Unlike traditional centralized methods, IPPO allows each agent to learn independently, reducing mutual interference during the learning process while capturing the characteristics of heterogeneous agents. In the IPPO algorithm, the policy network and value network of each agent are first initialized. The policy network is used to generate the agent's actions, and the value network is used to evaluate the value of the current state. Specifically, the parameters of the policy network... and parameters of the value network Initialized to random values ​​to begin the training process. During training, the IPPO algorithm runs the current policy of each agent in the environment. To collect trajectory data. Trajectory data includes observations at each time step. ,action ,award and the observation at the next moment This data is stored in an experience replay buffer. Subsequently, this data is used to calculate the advantage function to evaluate the merits of each action relative to the current policy. Advantage Function The calculation uses the generalized dominance estimation method, and its formula is:

[0183]

[0184] in, It is a discount factor that controls the impact of future rewards; V is the reward obtained by performing action i at time l; V(·) is the expected cumulative reward, T is the total number of time steps, and t is a single time step in T. The GAE method achieves a good balance between bias and variance, thus providing stable advantage estimation and effectively improving the performance of policy optimization. In the policy update phase, the IPPO algorithm updates the policy network parameters by optimizing the objective function. The objective function is as follows:

[0185]

[0186] in, This represents the probability ratio between the current strategy and the old strategy. This is a pruning parameter used to limit the magnitude of policy updates, preventing instability caused by excessively large policy update magnitudes. This represents the joint action of all agents. Through this objective function, the IPPO algorithm can effectively control the magnitude of policy updates, maintain training stability, and avoid excessive policy changes.

[0187] Furthermore, updating the value network is equally important. The objective function of the value network... Optimization is achieved by minimizing the error between the predicted value and the actual return, as shown in the following formula:

[0188]

[0189] The optimization process of the value network uses gradient descent to update the parameters. This is to reduce the error between the predicted value and the actual return.

[0190] Furthermore, the proposed dual-channel heterogeneous agent architecture comprises two sub-agents: an irrigation timing decision-making agent, which focuses on meteorological factors, soil moisture, and growth stages to determine when to irrigate; and an irrigation amount decision-making agent, which considers available water resources and crop growth needs to determine the appropriate water amount, preventing over-irrigation or under-irrigation. At the end of each growing season, the crop yield and irrigation costs calculated by the optimized prediction model are used to evaluate seasonal profits and serve as reward signals for the reinforcement learning agent to update the policy network, achieving continuous iteration and improvement of irrigation decisions. Through the coupling of "crop simulation model-driven—reinforcement learning policy optimization," the environment, while ensuring realistic diversity in meteorological, soil, and management measures, effectively reflects the water supply and demand patterns during crop growth, providing a high-fidelity simulation and learning platform for the research of intelligent water-saving irrigation strategies.

[0191] As an optional implementation, this embodiment also provides an agricultural irrigation decision optimization system, including:

[0192] The data pre-acquisition module is used to collect data from the target irrigation area to obtain soil data, climate data, crop information, and crop data.

[0193] The simulation model building module is used to input the soil data, climate data, and crop information into the AquaCrop model to obtain a crop simulation model.

[0194] The simulation module is used to simulate the crop yield and irrigation cost of the target irrigation area using the crop simulation model, and obtain model prediction data.

[0195] The difference calculation module is used to calculate the difference between the model prediction data and the crop data to obtain an error dataset.

[0196] An auxiliary model building module is used to train the improved GRU network using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model.

[0197] The decision model construction module is used to construct the Dec-POMDP model; the Dec-POMDP model includes: an irrigation time decision channel and an irrigation amount decision channel;

[0198] The decision generation module is used to input the soil data, climate data, and crop information into the Dec-POMDP model to generate irrigation decisions, resulting in irrigation time and irrigation amount decisions.

[0199] The decision result prediction module is used to input the irrigation time decision and the irrigation amount decision into the optimized prediction model to make predictions and obtain predicted yield and predicted cost.

[0200] The decision optimization module is used to iteratively optimize the Dec-POMDP model using the IPPO algorithm based on the predicted yield, the predicted cost, the crop yield, and the irrigation cost, to obtain the irrigation time optimization decision and the irrigation amount optimization decision.

[0201] The decision implementation module is used to perform irrigation treatment on the target irrigation area based on the irrigation time optimization decision and the irrigation amount optimization decision.

[0202] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to execute the aforementioned method for formulating qualification standards for operation ticket personnel based on multivariate knowledge analysis.

[0203] As an optional implementation, this embodiment also provides a non-transitory computer-readable storage medium storing computer instructions, which are used to cause a computer to execute the aforementioned method for formulating qualification standards for operation ticket personnel based on multivariate knowledge analysis.

[0204] The beneficial effects of this invention are as follows:

[0205] This invention improves the simulation effect and reliability of irrigation decisions by optimizing simulation results using an auxiliary model trained with the difference between the AquaCrop model and actual data. The constructed Dec-POMDP model flexibly responds to dynamic changes in environmental conditions and crop growth stages, and the two-stage decision-making improves the model's processing efficiency and reduces the consumption of water and computing resources.

[0206] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0207] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An agricultural irrigation decision optimization method, characterized in that, include: Data was collected from the target irrigation area to obtain soil data, climate data, crop information, and crop data. The soil data, climate data, and crop information are input into the AquaCrop model to obtain a crop simulation model; The crop simulation model is used to simulate the crop yield and irrigation cost of the target irrigation area to obtain model prediction data; Calculate the difference between the model's predicted data and the crop data to obtain an error dataset; The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model. Construct a Dec-POMDP model; the Dec-POMDP model includes: an irrigation time decision channel and an irrigation quantity decision channel; The soil data, climate data, and crop information are input into the Dec-POMDP model to generate irrigation decisions, resulting in irrigation time and irrigation amount decisions. The irrigation time decision and the irrigation amount decision are input into the optimized prediction model for prediction, and the predicted yield and predicted cost are obtained. Based on the predicted yield, the predicted cost, the crop yield, and the irrigation cost, the Dec-POMDP model is iteratively optimized using the IPPO algorithm to obtain irrigation time optimization decisions and irrigation amount optimization decisions; Irrigation treatment is carried out on the target irrigation area based on the irrigation time optimization decision and the irrigation volume optimization decision.

2. The agricultural irrigation decision optimization method according to claim 1, characterized in that, Data was collected from the target irrigation area to obtain soil data, climate data, crop information, and crop data, including: The soil data are obtained by collecting the total available soil water, soil water deficit, relative water deficit in the root zone, and basic soil texture parameters of the target irrigation area. The basic soil texture parameters include: soil type and porosity. The climate data is obtained by collecting rainfall data, crop evapotranspiration data, temperature, and relevant meteorological parameters in the target irrigation area; the relevant meteorological parameters include: relative humidity, wind speed, solar radiation, and CO2 concentration. The crop type, crop variety, and planting parameters of the target irrigation area are collected to obtain the crop information; The crop data are obtained by statistically analyzing the canopy coverage, aboveground biomass, growth stage, total water consumption, and seasonal water surplus of crops within the target irrigation area.

3. The agricultural irrigation decision optimization method according to claim 2, characterized in that, Data was collected from the target irrigation area to obtain soil data, climate data, crop information, and crop data, including: Outlier removal and missing value filling are performed sequentially on the soil data, the climate data, the crop information, and the crop data; Convert the soil type, crop type, and crop variety into unique thermal codes; The porosity, air temperature, and aboveground biomass were pretreated using the min-max normalization method. The Z-score normalization method was used to analyze the total available soil water, the soil water deficit, the rainfall data, the crop evapotranspiration data, and the... The concentration, wind speed, total water consumption, and seasonal remaining water volume are pretreated. The solar radiation and canopy coverage were preprocessed using the max normalization method. The relative water deficit in the root zone and the relative humidity are mapped proportionally to the unit interval.

4. The agricultural irrigation decision optimization method according to claim 3, characterized in that, The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including: The improved GRU network is constructed as follows: the improved GRU network includes an input layer, an encoding layer, and an output layer connected in sequence; the input layer includes an embedding layer, a first fully connected layer, a linear projection layer, and a feature vector fusion layer; the encoding layer includes a first improved GRU with embedded feature adaptive gating, a second improved GRU with bidirectional stacking, a temporal attention layer, a feature attention layer, a residual connection layer, a LayerNorm normalization layer, a Dropout layer, and a second fully connected layer connected in sequence; the embedding layer is connected to the first fully connected layer; the first fully connected layer and the linear projection layer are respectively connected to the feature vector fusion layer; The processed soil data, climate data, and crop information are categorized into static data and dynamic data. The static data is transformed using the embedding layer and the first fully connected layer to obtain a static vector.

5. The agricultural irrigation decision optimization method according to claim 4, characterized in that, The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including: The dynamic data is linearly mapped in time using the linear projection layer to obtain a dynamic vector, and the static vector and the dynamic vector are fused in time steps using the feature vector fusion layer to obtain an input sequence; The first improved GRU is used to extract features from the input sequence to obtain the hidden state; The second improved GRU is used to extract forward and backward dependencies from the hidden state to obtain bidirectional features; The temporal attention layer is used to calculate the time step weights and normalize the bidirectional features to obtain global temporal features; The feature attention layer is used to calculate the feature contribution weights of the global temporal features to obtain the key feature vector. The residual connection layer is used to add the key feature vector and the residual output of the second improved GRU to obtain the gradient reduction vector; The gradient reduction vector is normalized and randomly discarded using the LayerNorm normalization layer and Dropout layer, respectively, to obtain regularized features; The output difference branch of the output layer is used to perform dimensionality reduction mapping on the regularized features to obtain the predicted output difference; The regularized features are dimensionality-reduced and mapped using the cost difference branch of the output layer to obtain the predicted cost difference.

6. The agricultural irrigation decision optimization method according to claim 5, characterized in that, The improved GRU network is trained using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model, including: Construct the production error loss; the expression for the production error loss is: ;in, This refers to the production error loss; This represents the total time step; Weights for time-series stages; Huber loss function; For model predictions Time step output difference; for The difference between the actual output at each time step; Construct cost error loss; the expression for the cost error loss is: ;in, This refers to the cost error loss; Cost weighting; for The soil moisture deficit at the time step; For model predictions Time step cost difference; for The actual cost difference of each time step; Construct a regular expression; the expression for the regular expression is: ;in, For the regularization term; The number of samples; , The first Predicted values ​​of production difference and cost difference for each sample; , These are the mean error of historical output difference and the mean error of historical cost difference, respectively. Integrating the production error loss, the cost error loss, and the regularization term, we obtain the overall loss function; the expression for the overall loss function is: ;in, The overall loss function; This is the dynamic balance coefficient; The regularization coefficient is used. The improved GRU network is iterated using the overall loss function based on the predicted cost difference and the error dataset to obtain the optimized prediction model.

7. The agricultural irrigation decision optimization method according to claim 6, characterized in that, Constructing the Dec-POMDP model includes: Construct a state space; the state space includes: , Month, DAP, IrrCum, IrrSur, CC, B, GrowthStage, Dep, TAW, Depletion, Rain1, Rain2, Rain3, ET1, ET2, ET3; among them, Month, DAP, IrrCum, IrrSur, CC, B, GrowthStage, Dep, TAW, Depletion, Rain1, Rain2, Rain3, ET1, ET2, and ET3 represent the date, month, total number of days since planting, total water consumption, seasonal remaining water, canopy coverage, aboveground biomass, growth stage, relative water deficit in the root zone, total available soil water, soil water deficit, average daily rainfall over the past 7 days, total rainfall, yesterday's rainfall, average daily reference crop evapotranspiration over the past 7 days, total crop evapotranspiration, and yesterday's reference crop evapotranspiration. Determine the number of agents; the number of agents is 2. Construct an action space; the expression of the action space includes: and ;in, To select any day from day 0 to day 13 after the current decision as the irrigation time; For the selected irrigation amount; Construct an observation space; the expression for the observation space is: and ;in, This refers to the observation space of the irrigation time decision channel; This refers to the observation space of the irrigation volume decision channel; Define the state transition probability; Construct a reward function; the expression of the reward function is: ;in, The reward function; For time step The state of being; For time step Joint actions at the location; For crop prices; The unit cost of irrigation; The total yield of the crop on the harvest day; For the first Daily irrigation volume; The fixed non-irrigated production cost for this round at the end of the simulation; Define a discount factor; the expression for the discount factor is: ;in, Let the objective function be the policy objective function; For corresponding actions and state Expectations; The discount factor; For instant rewards; The Dec-POMDP model is obtained by integrating the state space, the number of agents, the action space, the observation space, the state transition probability, the reward function, and the discount factor.

8. An agricultural irrigation decision optimization system, characterized in that, include: The data pre-acquisition module is used to collect data from the target irrigation area to obtain soil data, climate data, crop information, and crop data. The simulation model building module is used to input the soil data, climate data, and crop information into the AquaCrop model to obtain a crop simulation model. The simulation module is used to simulate the crop yield and irrigation cost of the target irrigation area using the crop simulation model, and obtain model prediction data. The difference calculation module is used to calculate the difference between the model prediction data and the crop data to obtain an error dataset. An auxiliary model building module is used to train the improved GRU network using the error dataset to obtain a trained auxiliary calibration model. The crop simulation model and the auxiliary calibration model are then integrated to obtain an optimized prediction model. The decision model construction module is used to construct the Dec-POMDP model; the Dec-POMDP model includes: an irrigation time decision channel and an irrigation amount decision channel; The decision generation module is used to input the soil data, climate data, and crop information into the Dec-POMDP model to generate irrigation decisions, resulting in irrigation time and irrigation amount decisions. The decision result prediction module is used to input the irrigation time decision and the irrigation amount decision into the optimized prediction model to make predictions and obtain predicted yield and predicted cost. The decision optimization module is used to iteratively optimize the Dec-POMDP model using the IPPO algorithm based on the predicted yield, the predicted cost, the crop yield, and the irrigation cost, to obtain the irrigation time optimization decision and the irrigation amount optimization decision. The decision implementation module is used to perform irrigation treatment on the target irrigation area based on the irrigation time optimization decision and the irrigation amount optimization decision.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the method for formulating qualification standards for operation ticket personnel based on multivariate knowledge analysis as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute any one of claims 1 to 7, a method for formulating qualification standards for operation ticket personnel based on multivariate knowledge analysis.

Images