Coastal saline soil drainage and salinity intelligent regulation method and system based on artificial intelligence

Abstract

The present application relates to the technical field of soil drainage and salinity improvement, and in particular to an intelligent control method and system for coastal saline-alkali soil drainage and salinity improvement based on artificial intelligence, which specifically comprises the following steps: first, dividing a treatment unit and laying out sensing devices, collecting multi-source heterogeneous data to form a data set; cleaning the data and synchronizing in time and space, then using segmented physical perception standardization processing to obtain a standardized feature vector; constructing a graph neural network micro-zone irrigation demand prediction model to predict and recommend irrigation amount and soil salinity change amount; calculating the loss function of the model and iteratively training the model; periodically performing decision-making control cycles on the trained model, converting the output of the multi-source heterogeneous data collected at each decision-making time into instructions and synchronously performing drainage control. The present application can realize optimal control of water and salt environment and form an intelligent treatment closed loop.

Description

Artificial Intelligence-Based Intelligent Regulation Method and System for Drainage and Salt Improvement in Coastal Saline-Alkali Land Technical Field

[0001] This invention relates to the field of soil drainage and salinization technology, and in particular to an intelligent regulation method and system for drainage and salinization of coastal saline-alkali land based on artificial intelligence. Background Technology

[0002] Saline-alkali land improvement is an important technical direction for improving soil quality and ensuring agricultural production. Although the traditional treatment model has established a technical paradigm with "irrigation and leaching" and "engineering drainage" as the core, in actual application, these two methods are often in a state of separation. They rely on fixed procedures or experience-based scheduling and lack dynamic coordination based on real-time system status. This results in problems such as incomplete improvement, easy salinization, water waste, and lack of systemic regulation, making it difficult to achieve integrated intelligent treatment.

[0003] A search revealed existing technologies for intelligent control systems for saline-alkali land, such as the invention patent with publication number CN115081317A, which proposes a method for predicting the dynamic changes of soil salinization parameters in saline-alkali land. However, these existing technologies objectively have the following shortcomings:

[0004] Conventional techniques typically perform simple time-series alignment and mathematical interpolation on independent data sources, failing to fully consider the physical correlations and scale effects between micro-area data and vertical shaft data, and between point-scale measurements and area-scale meteorological data. This results in a fused data sequence that, while aligned at specific points in time, cannot constitute a physically self-consistent snapshot of the system's state, thus affecting the accuracy of subsequent analyses. Furthermore, conventional standardization or normalization methods rely purely on linear transformations based on the statistical distribution of the data. This processing can smooth out or distort the inherent relationships between different characteristics (such as irrigation volume, salinity, and water level) determined by physical laws such as water-salt balance and mass conservation. The input model's feature set itself suffers from physical information loss, potentially misleading the model into learning correlations that violate basic scientific principles. Existing regression models or independent neural networks based on single-point data generally treat each treatment unit or monitoring point as an isolated entity for modeling and prediction, failing to fully consider the spatial correlations and overall system coupling brought about by lateral transport of soil water and salt, groundwater connectivity, and shared drainage facilities. In conventional saline-alkali land treatment, drainage and irrigation leaching are often managed as two independent processes, based on historical experience or fixed procedures, lacking a dynamic collaborative scheduling mechanism based on real-time system status and future operation predictions. This fragmented approach cannot proactively and forward-lookingly adjust drainage intensity during leaching to suppress water level rise, easily leading to the dilemma of "treating while salinizing," resulting in low treatment efficiency and high energy consumption.

[0005] Therefore, this invention proposes an intelligent regulation method and system for drainage and salt conversion in coastal saline-alkali land based on artificial intelligence to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes an intelligent regulation method and system for drainage and salt conversion in coastal saline-alkali land based on artificial intelligence. By integrating multi-source heterogeneous data with high consistency and accurately characterizing the transient state of the system, this invention enables model predictions to conform to physical laws and have spatial coordination, thereby facilitating efficient drainage and salt conversion.

[0007] On the one hand, the technical solution of this invention to solve the technical problem is an intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence, which includes the following steps:

[0008] S1. Divide the treatment unit into treatment units, install shafts and drainage devices in each treatment unit, and set up the first sensor group in the shaft to collect groundwater level and groundwater quality data; divide each treatment unit into multiple leaching micro-zones, set up the second sensor group in each leaching micro-zone to collect soil moisture information; also collect external meteorological data and agricultural record data to form a dataset.

[0009] S2. For the multi-source heterogeneous data in the dataset, clean and synchronize the data in time and space based on the monitoring time of the shaft to obtain the feature values ​​after cleaning and synchronization, and form a cleaned and synchronized dataset.

[0010] S3. For the data in the cleaning synchronization dataset, a segmented physical perception standardization method is adopted. The data is classified according to the physical meaning of the features. The appropriate standardization transformation is applied to different types of features. The water-salt balance residual term is introduced to modulate the intensity characteristic features obtained by classification. The standardized features of various types are integrated to obtain the standardized feature vector.

[0011] S4. Construct a micro-area irrigation demand prediction model based on graph neural network. Input standardized feature vectors, use standardized feature vectors as nodes to construct physical perception features, input the initial hidden state features containing physical perception features into the physical-guided graph neural network model for encoding, output deep encoded features, and then decode through a two-branch process to output the predicted value of recommended irrigation amount and the change in soil salinity.

[0012] S5. Calculate the total loss function of the model and train the model based on the total loss function;

[0013] S6. The trained model performs periodic decision-making and control cycles, converting the multi-source heterogeneous data collected at each decision moment into instructions after passing through the model and simultaneously performing drainage control.

[0014] S1 is as follows:

[0015] The land parcels were divided into several independent governance units to form distributed coastal saline-alkali land.

[0016] Vertical plastic paving is carried out around each treatment unit to form a vertical seepage barrier wall;

[0017] The second sensor group installed in the rinsing micro-zone includes a soil temperature and humidity sensor, a soil conductivity sensor, a soil salinity sensor, and a soil pH sensor; the rinsing pipeline network is equipped with a rinsing solenoid valve in each rinsing micro-zone to control the rinsing independently;

[0018] Vertical shafts and drainage devices are installed in each treatment unit. Specifically, underground pipes with filter media are laid below the cultivated layer in each treatment unit. Vertical shafts are drilled and connected to the underground pipes. Drainage ditches are excavated on the surface in the treatment unit. Variable frequency submersible pumps are installed in the vertical shafts. The submersible pumps are connected to drainage pipes, and the end of the drainage pipes is led to the drainage ditch.

[0019] The dataset includes soil moisture information, groundwater level data, groundwater quality data, external meteorological data, and agricultural records.

[0020] S2 is as follows:

[0021] S2.1, Using the stable and low-frequency monitoring times of the sensor group installed in the vertical shaft as the time alignment benchmark, through parameters... This represents the number of valid data acquisition moments of the sensor array in the shaft within the observation period, forming the reference time axis. , ;

[0022] in, Represents the first on the reference time axis Each data collection moment, ;

[0023] S2.2 For any reference time, synchronize the sensor data of each leaching micro-zone to which it belongs. Specifically, for the soil moisture information collected by the micro-zone sensors, use an adaptive time window weighted average based on the soil temperature change rate to interpolate and calculate the synchronized soil attribute values ​​of each leaching micro-zone at each collection time on the reference time axis.

[0024] Soil moisture information includes relatively slow-changing attributes such as soil volumetric water content, soil temperature, electrical conductivity, salinity, and pH.

[0025] S2.3 External meteorological data are macro-meteorological data from external data sources, including raw temperature and raw recent cumulative evaporation, which are aligned using the nearest neighbor principle;

[0026] Agricultural record data consists of discrete records or periodic statistical values, including crop growth period codes, raw recent cumulative irrigation volume, and raw cumulative working hours recorded by vertical well submersible pumps. Based on the time attributes of the agricultural record data, it is directly associated with the specific date or statistical period of each benchmark time, and the corresponding values ​​are extracted and aligned.

[0027] S2.4 For each reference time and each rinsing micro-zone in the shaft monitoring, generate a multidimensional data vector containing all aligned features, and integrate the multidimensional data vectors of all times and all micro-zones to form a cleaning synchronization dataset.

[0028] S3 is as follows:

[0029] S3.1 Divide the cleaning synchronization dataset into intensity characteristic features, cumulative value features, and state coding features according to the physical meaning and data properties of the features;

[0030] S3.2. Physical perception standardization is performed on the strength characteristics. The residual term based on the simplified water-salt balance equation is used to modulate the traditional Z-score standardization process to obtain the standardized strength characteristics.

[0031] S3.3. Perform logarithmic scaling standardization on the cumulative features. First, perform logarithmic scaling to alleviate data skew, and then perform Z-score standardization to obtain the standardized cumulative features.

[0032] S3.4. For state coding features, one-hot coding is used to convert them into binary vectors;

[0033] S3.5 Integrate the standardized intensity characteristics and cumulative characteristics, as well as the binary vector, to obtain the standardized feature vector of each rinsing micro-region at each reference time.

[0034] S4 is as follows:

[0035] S4.1. Using the standardized feature vector as the basic feature of each rinsing micro-zone node, two additional physical perception features are extracted and constructed from the cleaning synchronization dataset and spliced ​​with the basic features to form the initial hidden state feature vector of each node.

[0036] S4.2, Graph Neural Network Models include A stacked graph attention network layer is used to aggregate information about each node itself and its first-order neighbors. Layer, for the target node With any of its neighboring nodes The original attention score is calculated by comparing the physical state with the introduced guiding signal.

[0037] Based on normalized attention weights, nodes In the The updated features of the layer are obtained by weighted aggregation of the transformed features of neighboring nodes and nonlinear activation to obtain the updated hidden state feature vector.

[0038] go through After layer encoding, we obtain deep encoded features for each node that contain information about its higher-order neighbors.

[0039] S4.3. A dual-branch decoding structure is adopted to collaboratively predict and verify deep coding features. The first branch predicts the recommended irrigation amount in the future decision-making cycle through the first multilayer perceptron, and the second branch predicts the expected change in soil salinity at future times through the second multilayer perceptron.

[0040] S5 is detailed below:

[0041] S5.1 First, based on the prediction results of the micro-area irrigation demand prediction model and the actual labels, calculate the irrigation volume prediction loss term and the water and salt status prediction loss term. The actual labels are the actual irrigation volume and the actual water and salt status. Then, based on the model prediction results, calculate the physical consistency constraint loss term, and introduce the spatial smoothing regularization term and the theoretical residual equation. The spatial smoothing regularization term is a regularization term that encourages the smoothing of irrigation decisions in adjacent micro-areas. The difference between the predicted salinity change and the theoretical salinity change is calculated through the theoretical residual equation. Finally, the three loss terms are weighted and added together to obtain the total loss function.

[0042] S5.2, Iteratively train the model;

[0043] The dataset is divided into training, validation, and test sets according to time. The model parameters are initialized with a standard neural network. In each training round, a batch of samples is randomly selected, and the prediction results are output through forward propagation. After calculating the total loss of the batch, the gradient is calculated by backpropagation. The parameters are updated using an adaptive optimization algorithm to minimize the loss. One iteration of the training set is considered one cycle. At the end of each cycle, the validation set is used for evaluation. A stopping condition is set. After training stops, the optimal parameter model of the validation set is saved, and its generalization ability is evaluated using the test set.

[0044] S6 is detailed below:

[0045] The trained model periodically executes the following decision-making and control loop;

[0046] First, multi-source heterogeneous data is automatically collected at each decision moment. Then, the data is cleaned, spatiotemporally synchronized, and physically standardized to generate standardized feature vectors for all rinsing micro-regions at the current moment.

[0047] Then, the standardized feature vector, the constructed physical perception features, and the graph structure defined according to the governance unit layout are input into the trained micro-area irrigation demand prediction model. Each leaching micro-area outputs a recommended irrigation demand prediction value for a future decision cycle.

[0048] Based on the predicted irrigation demand values ​​of each micro-region, specific execution instructions are converted, recommended irrigation amounts are issued for each micro-region, and drainage regulation is carried out simultaneously.

[0049] On the other hand, the present invention also provides an intelligent regulation and control system for drainage and salt conversion of coastal saline-alkali land based on artificial intelligence, including a module for executing processing instructions for each step in an intelligent regulation and control method for drainage and salt conversion of coastal saline-alkali land based on artificial intelligence, as follows:

[0050] Zoning isolation module: includes a geomembrane laid vertically along the boundary of the treatment unit to block lateral saline water intrusion;

[0051] Subsurface drainage module: includes subsurface pipes laid below the topsoil layer for draining saline water from the soil after leaching;

[0052] Vertical shaft drainage module: includes a vertical shaft, which is equipped with a drainage device, which includes a variable frequency submersible pump and a drainage pipe, for collecting the rinsing salt water discharged from the underground pipe and pumping it efficiently to the external drainage ditch; the end of the underground pipe is connected to the vertical shaft, the variable frequency submersible pump is installed in the vertical shaft, the submersible pump is connected to the drainage pipe, and the end of the drainage pipe is led to the drainage ditch.

[0053] Irrigation and leaching module: used to extract fresh water from irrigation ditches to leach saline-alkali land; the irrigation and leaching module includes a leaching pipe network, which is connected to an irrigation pump, and a leaching solenoid valve is installed on each leaching micro-zone on the leaching pipe network.

[0054] Monitoring module: includes a first sensor group and a second sensor group, used to monitor soil moisture, water level and water quality data;

[0055] Intelligent collaborative control module: including controller, which is connected to the vertical shaft drainage module, irrigation and rinsing module and monitoring module respectively, and is used to dynamically control rinsing and drainage;

[0056] Wind-solar hybrid power generation system: It is electrically connected to each power-consuming module to supply power.

[0057] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0058] This invention discloses an intelligent regulation method and system for drainage and salt conversion in coastal saline-alkali land based on artificial intelligence. It adopts a spatiotemporal synchronous fusion method of multi-source heterogeneous data based on physical process constraints, abandons the conventional independent time alignment and interpolation techniques, and creatively uses the monitoring time of the vertical shaft, which reflects the overall state of the system, as the reference time axis. It also designs an adaptive time window weighted average interpolation algorithm that combines the soil temperature change rate for micro-area sensor data, ensuring that data from different spatial scales and acquisition frequencies can achieve high consistency fusion at a unified physical time, and accurately characterize the transient real state of the "micro-area-vertical shaft" coupled system.

[0059] This invention incorporates a perceptual data standardization strategy based on the physical laws of water-salt balance. To address the problem that conventional standardization methods can destroy the physical correlation between features, it proposes a segmented physical perceptual standardization. Its core lies in introducing a residual term calculated based on a simplified water-salt balance equation as a modulation factor during the standardization process of the "intensity characteristic" feature. This ensures that the standardized data not only has unified dimensions, but its numerical value can also directly reflect the degree to which the current state of the system deviates from the water-salt balance, providing a feature expression with embedded physical meaning for subsequent models.

[0060] This invention also employs a physically-guided spatiotemporal graph neural network prediction model architecture. To model the spatial correlation and system coupling of micro-intervals, the governance unit is abstracted into a dynamic heterogeneous graph, and three key physical enhancements are implemented: First, the "historical water-salt pressure index" and "drainage potential coefficient" are integrated into the initial features of nodes; second, in the graph attention mechanism, in addition to node feature similarity, "water-salt balance residual similarity" is introduced as a physical guidance signal, directly driving information to preferentially aggregate among nodes with similar physical states; third, a dual-branch decoder is used to simultaneously predict irrigation demand and salinity changes, forming an inherent physical verification loop.

[0061] It also integrates a composite model training strategy that combines physical constraints and spatial regularization, and adopts a composite loss function that goes beyond the conventional mean square error. This function not only includes the prediction errors of irrigation amount and salinity changes, but also innovatively adds a "physical consistency constraint loss" term. By simplifying the theoretical equation, the predicted irrigation amount is mapped to the theoretical salinity change, and the deviation of the model's predicted change from this theoretical value is penalized. At the same time, a spatial regularization term that encourages smooth decision-making in neighboring micro-regions is added, thereby forcing the mapping relationship learned by the model to satisfy basic physical laws and have spatial consistency. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0063] Figure 1 is a flowchart illustrating the method of the present invention.

[0064] Figure 2 is a flowchart of step S2.

[0065] Figure 3 is a flowchart of step S3.

[0066] Figure 4 is a flowchart of step S4.

[0067] Figure 5 is a flowchart of step S5.

[0068] Figure 6 is a schematic diagram of the structure of saline-alkali land management divided into multiple management units.

[0069] Figure 7 is a schematic diagram of the structure of a single governance unit.

[0070] Figure 8 is a cross-sectional view of the connection between the underground pipe, the vertical shaft and the drainage ditch.

[0071] Figure 9 is a three-dimensional view of the vertical seepage barrier wall formed after the treatment unit is vertically paved with plastic.

[0072] Figure 10 is a schematic diagram of a single treatment unit divided into multiple rinsing micro-zones.

[0073] Figure 11 is a schematic diagram of a wind-solar hybrid power generation system.

[0074] Figure 12 is a control block diagram of the present invention.

[0075] Figure 13 is a schematic diagram showing the performance of the micro-area irrigation demand prediction model based on graph neural network in predicting micro-area irrigation demand.

[0076] Figure 14 is a schematic diagram illustrating the performance of the fully connected neural network model in predicting irrigation demand in micro-areas.

[0077] Figure 15 is a schematic diagram illustrating the performance of the random forest regression model in predicting irrigation demand in micro-areas.

[0078] Figure 16 is a schematic diagram illustrating the performance of the linear regression model in predicting irrigation demand in micro-areas.

[0079] Explanation of reference numerals in the attached figures:

[0080] 1. Treatment unit; 2. Vertical anti-seepage wall; 3. Concealed pipe; 4. Flushing pipe network; 5. Flushing solenoid valve; 6. Shaft; 7. Drainage pipe; 8. Drainage ditch; 9. Irrigation ditch; 10. Irrigation pump; 11. Wind turbine; 12. Photovoltaic panel; 13. Variable frequency submersible pump; 14. Wind-solar hybrid controller; 15. Storage battery; 16. Inverter. Detailed Implementation

[0081] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, the components and arrangements of specific examples are described below.

[0082] Example 1

[0083] As shown in Figure 1, an intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence includes the following steps:

[0084] S1. Divide the treatment unit into treatment units, install shafts and drainage devices in each treatment unit, and set up the first sensor group in the shaft to collect groundwater level and groundwater quality data; divide each treatment unit into multiple leaching micro-zones, set up the second sensor group in each leaching micro-zone to collect soil moisture information; also collect external meteorological data and agricultural record data to form a dataset.

[0085] S2. For the multi-source heterogeneous data in the dataset, clean and synchronize the data in time and space based on the monitoring time of the shaft to obtain the feature values ​​after cleaning and synchronization, and form a cleaned and synchronized dataset.

[0086] S3. For the data in the cleaning synchronization dataset, a segmented physical perception standardization method is adopted. The data is classified according to the physical meaning of the features. The appropriate standardization transformation is applied to different types of features. The water-salt balance residual term is introduced to modulate the intensity characteristic features obtained by classification. The standardized features of various types are integrated to obtain the standardized feature vector.

[0087] S4. Construct a micro-area irrigation demand prediction model based on graph neural network. Input standardized feature vectors, use standardized feature vectors as nodes to construct physical perception features, input the initial hidden state features containing physical perception features into the physical-guided graph neural network model for encoding, output deep encoded features, and then decode through a two-branch process to output the predicted value of recommended irrigation amount and the change in soil salinity.

[0088] S5. Calculate the total loss function of the model and train the model based on the total loss function;

[0089] S6. The trained model performs periodic decision-making and control cycles, converting the multi-source heterogeneous data collected at each decision moment into instructions after passing through the model and simultaneously performing drainage control.

[0090] In a specific implementation, S1 is as follows:

[0091] The site was divided into several independent treatment units through investigation. Vertical plastic paving was carried out around each treatment unit to form a vertical anti-seepage wall. Each treatment unit was further divided into multiple rinsing micro-zones.

[0092] Below the topsoil of each treatment unit, underground pipes with filter media are laid at designed intervals. Vertical shafts are drilled and connected to the underground pipes. Finally, drainage ditches are excavated on the surface of the treatment unit. Variable frequency submersible pumps are installed in the vertical shafts, and the submersible pumps are connected to drainage pipes, the ends of which lead to the drainage ditches. The first sensor group is installed in the vertical shafts, including, for example, groundwater level sensors and groundwater quality sensors: the groundwater level sensor is used to track the water level changes in real time after drainage (in the vertical shafts), ensuring that it is stably maintained below the "critical depth for salinization"; the groundwater quality sensor is used to monitor the decrease in groundwater salinity during drainage, and combined with soil conductivity / salt data, to determine the effectiveness of the desalination operation.

[0093] A second sensor group is installed in each rinsing micro-zone. The sensor group is connected to the controller through control lines. The sensor group includes, but is not limited to, soil temperature and humidity sensors, soil conductivity sensors, soil salinity sensors, and soil pH sensors. At the same time, a wind-solar hybrid power supply module is installed at a suitable location in the treatment unit for power supply.

[0094] A rinsing pipe network is laid in the treatment unit, and the rinsing pipe network is connected to the irrigation pump. A rinsing solenoid valve is installed on each rinsing micro-zone on the rinsing pipe network to realize independent rinsing control.

[0095] In a specific implementation, S2 is as follows:

[0096] Because the sensor data, external meteorological data, and agricultural records of the leaching micro-area and its associated shafts are multi-source, heterogeneous, and not fully synchronized in terms of acquisition time, the micro-area sensors collect soil physicochemical data at a higher frequency, the shaft sensors monitor groundwater level and water quality at a lower frequency, the meteorological data comes from external stations at a macro scale, and the agricultural records are discrete event records. Conventional techniques usually perform independent time alignment and interpolation processing on each data source, ignoring the physical correlation and spatial scale differences between the micro-area, shaft, and meteorological data. This results in poor physical consistency after data fusion, making it impossible to accurately represent the true state of the "micro-area-shaft" system at a unified moment.

[0097] Therefore, this invention adopts a synchronous cleaning method based on the monitoring time of the shaft and combined with the spatial weight constraint of the physical process to achieve highly consistent fusion of multi-source data at the same physical moment, as shown in Figure 2. The specific steps are as follows:

[0098] S2.1. Using the stable, low-frequency data acquisition time sequence of the shaft sensor as the benchmark for time alignment of the entire system, assume that the shaft sensor has a total of [number missing] data acquisition time sequences within the observation period. The effective acquisition times constitute the reference time axis. ;

[0099] in, The reference time axis represents the collection of all valid acquisition moments from the vertical shaft sensors, serving as a reference for time synchronization of the entire system. Represents the first on the reference time axis Each data collection moment, Represents the first on the reference time axis Each data collection moment; Represents the time index, with a value range of 100. ; This represents the total number of valid acquisition moments of the shaft sensor throughout the entire observation period, i.e., the length of the reference time axis, and is a positive integer.

[0100] S2.2. For any reference time, synchronize the sensor data of each leaching micro-zone to which it belongs. Specifically, for the relatively slowly changing attribute data of soil volumetric water content, electrical conductivity, salinity, soil temperature, and pH collected by the micro-zone sensors, use an adaptive time window weighted average based on the soil temperature change rate for interpolation calculation to reflect the physical inertia of the soil process, expressed as:

[0101]

[0102] In the formula, This indicates the synchronized soil attribute value, which is the first... Each shower micro-zone at the reference time Soil property data obtained after cleaning and synchronization processing is soil property data obtained after cleaning and synchronization processing. This represents a micro-area index, used to identify the wash micro-area number to which the data belongs, for example... Indicates the first Micro-regions; This represents the initial data acquisition time of the micro-area sensor, which is the time window. A point in time within; Represented by reference time The time window centered on this point is used to select the original time moments for interpolation calculations, and the width of the time window is [value missing]. ; Indicates the assignment of time. The adaptive weights for the data incorporate soil temperature variation information in their calculation to reflect the dynamic characteristics of the physical process, and are expressed as follows: ; Indicates the first Each micro-region at time The original feature values ​​obtained from the measurement at the location; Represents the natural exponential function; This represents the temperature change sensitivity coefficient, a preset hyperparameter used to adjust the intensity of the influence of the soil temperature change rate on the weight decay rate. An example value is shown below. Indicates at time No. Original soil temperature of each micro-region The absolute value of the rate of change over time is specifically approximated by using first-order discrete difference to calculate the original temperature data, representing... The degree of drastic change in soil temperature at any given time; Indicates the first The original soil temperature of each micro-region is the first The raw soil temperature measurement value of each micro-area is directly collected by the sensor, and the unit is degrees Celsius.

[0103] In practical implementation, the width of the time window Examples of possible values ​​are: The time was measured for hours to ensure that enough original micro-area sampling points were covered for weighting. The original characteristics obtained from the micro-area measurements included soil volumetric water content (in percentage, representing the volume ratio of water in the soil), soil electrical conductivity (in millisiemens per centimeter, used to indirectly reflect the salt concentration of the soil solution), soil salinity (i.e., the mass of soluble salts contained in a unit mass of soil, in grams per kilogram), soil temperature (in degrees Celsius), and soil pH (a dimensionless indicator that reflects the acidity or alkalinity of the soil solution).

[0104] It should be noted that when performing calculations, the soil volumetric water content must be converted to a decimal for calculation.

[0105] S2.3 For macro-meteorological data originating from external data sources, including raw temperature... and the original recent cumulative evaporation Alignment is performed using the nearest neighbor principle. Specifically, for each reference time, the meteorological observation or forecast value of the nearest available time before and after that time is directly selected as its synchronization value.

[0106] For discrete records or periodic statistics such as crop growth period codes, raw recent cumulative irrigation volume, and raw cumulative working hours recorded by vertical well submersible pumps, the corresponding values ​​are extracted by directly associating them with the specific date or statistical period of each baseline time based on the time attribute of the records.

[0107] definition: The crop growth stage code is a discrete status label that identifies the current growth stage of the crop. Indicates the first The original recent cumulative irrigation amount recorded in each micro-area; This indicates the original cumulative working time recorded by the submersible pump in the vertical shaft.

[0108] S2.4 After completing the alignment processing of all data sources, for each reference time and each wash micro-region, generate a multidimensional data vector containing all aligned features, and integrate the multidimensional data vectors of all time points and all micro-regions to form a washing synchronization dataset. ;

[0109] in, Represents a multidimensional data vector, containing the first... Each shower micro-zone at the reference time All feature values ​​after cleaning synchronization; This means "for all micro-regions". This means "for all reference times".

[0110] In a specific implementation, S3 is as follows:

[0111] The data features in the synchronized data set exhibit significant differences in dimensions and magnitudes, and there may be clear physical relationships between them. Conventional standardization methods, which only perform statistical linear transformations, can easily disrupt the inherent physical relationships between features, potentially leading the model to learn feature combinations that violate physical laws. Therefore, this invention employs a segmented physical-aware standardization method, classifying features based on their physical meaning and applying appropriate standardization transformations to different types of features. Specifically, for "intensity characteristic" features, water-salt balance residuals are introduced for modulation and transformation, ensuring that the standardized data, while maintaining uniform dimensions, reflects the degree to which the system deviates from the water-salt balance state, as shown in Figure 3. The specific steps are as follows:

[0112] S3.1. Based on the physical meaning and data properties of the features, clean the synchronous dataset. All features are divided into the following three categories:

[0113] (1) “Strength characteristics”

[0114] The "intensity characteristic" refers to the instantaneous or quasi-instantaneous measurement of the system state at a certain moment, denoted by the first... Taking a micro-region as an example, the "intensity characteristics" feature includes:

[0115] Indicates the first Soil volumetric water content in individual micro-regions Indicates the first Soil electrical conductivity in individual micro-regions Indicates the first Soil salinity in individual micro-regions Indicates the first Soil temperature in individual micro-regions Indicates the first Soil pH in individual micro-regions This indicates the groundwater depth monitored by the shaft (a characteristic shared by all micro-areas to which the shaft belongs). This indicates the groundwater salinity monitored by the vertical shaft (a characteristic shared by all micro-areas to which the vertical shaft belongs). It represents temperature (a characteristic shared by all micro-regions).

[0116] (2) Characteristics of "cumulative amount"

[0117] The "cumulative amount" characteristic represents the cumulative effect of events over a period of time, with the first... Taking a micro-region as an example, the "cumulative amount" characteristic includes:

[0118] This represents the recent cumulative evaporation (a characteristic shared by all microregions). Indicates the first Recent cumulative irrigation amount for each micro-area This indicates the recent cumulative operating time of the submersible pump in the shaft (a characteristic shared by all micro-areas to which the shaft belongs).

[0119] (3) "Status coding" feature

[0120] The "state coding" feature represents discrete system states, including: This indicates the crop growth period code.

[0121] S3.2, Physical Perception Standardization of "Strength Characteristics"

[0122] For the "intensity characteristics" feature, the traditional Z-score standardization process is modulated using the residual terms based on the simplified water-salt balance equation. Taking soil salinity content as an example, the standardization method is expressed as follows:

[0123]

[0124] In the formula, Indicating standardized soil salinity, it is the first The output value of the salinity characteristics of each micro-region after physical sensing standardization represents the soil salinity level relative to the average level of all samples under a unified dimension, and takes into account the current water and salt balance state of the system. This represents the overall arithmetic mean of the salt content in all training samples, used for data centralization; This represents the overall standard deviation of salt content across all training samples, used for data scaling. This represents the balance coefficient sensitivity, a preset hyperparameter used to adjust the intensity of the influence of the water-salt balance residuals on the normalization scaling factor. Example values ​​are provided. ; Indicates the first The simplified water-salt balance residuals of each micro-region, expressed in units of equivalent water flux intensity (e.g., mm / day), characterize the difference between water-salt input and output, and are used to quantify the micro-region. The imbalance between water and salt input and output at the current moment is calculated as follows: ; Indicates the first The cumulative irrigation amount of each micro-area within the recent statistical period. The recent statistical period is the statistical duration used to convert the cumulative amount into the average intensity; an example value is 7 days. At the reference time Total irrigation volume over the previous 7 days; This represents the salt leaching water demand coefficient, a preset empirical parameter used to convert soil salinity and water content into equivalent salt leaching water flux in the simplified water-salt balance equation. An example value is 0.1. This represents the length of the statistical period for the cumulative amount, in days. It is used to convert the cumulative amount into the average intensity within the period. An example of its value is shown below. sky; This represents the effective drainage intensity calculated from the cumulative operating time of the submersible pump in the vertical shaft. It not only considers the pump's operating time but also couples the impact of the current groundwater depth on drainage capacity through a drainage efficiency function. The calculation method is expressed as follows: ; A function representing the drainage efficiency related to the current groundwater depth; The leaching efficiency coefficient is a constant that represents the proportion of salt that can be leached away by a unit volume of irrigation water. Examples of its values ​​are as follows: .

[0125] In practical implementation, the remaining "intensity characteristics" adopt standardized formulas of similar form, and their modulation factors can be adjusted according to the physical meaning of the characteristics themselves, or directly reused from those shared with soil salinity content. value.

[0126] In practical implementation, the drainage efficiency function The specific form needs to be calibrated based on field observations or hydraulic models, and the value ranges from 0 to 1. For example, define... This characterizes the actual drainage efficiency of a submersible pump per unit working time at the current groundwater depth. The first empirical parameter is a negative value, for example... , The second empirical parameter is a positive value, for example... , This indicates the operation of retrieving the maximum value. This indicates the operation of taking the minimum value.

[0127] It should be noted that the simplified method for calculating the water-salt balance residual is a simplified equation. As a conceptual expression of the salinity term, its specific form can be determined based on the principle of salinity balance in practical applications, or its effective coefficient can be determined through data fitting. The term is an empirical term, which will It is redefined as a comprehensive empirical coefficient, with dimensions of "equivalent water flux intensity per unit salt mass concentration and per unit water content". In actual calculations... A numerical value is obtained through data calibration, which implicitly involves unit conversion and simplification of physical relationships.

[0128] S3.3 Logarithmic scaling normalization of the "cumulative" feature

[0129] For the "cumulative amount" feature, since its data distribution may exhibit a long-tailed shape, logarithmic scaling is first performed to alleviate data skew, followed by Z-score standardization. Taking the micro-area cumulative irrigation amount as an example, its standardization method is expressed as follows:

[0130]

[0131] In the formula, Represents standardized cumulative irrigation volume, which is the first The output value of the cumulative irrigation amount feature of each micro-region after logarithmic scaling and Z-score standardization represents the degree of deviation of the cumulative irrigation amount from the mean of the logarithmically transformed sample under a unified dimension. This represents a logarithmic function, with the default base being the natural constant. Indicates that among all training samples The overall arithmetic mean of the values; Indicates that among all training samples The overall standard deviation of the values.

[0132] In practice, all other "cumulative" features are processed using the same method of "logarithmic transformation followed by Z-score standardization".

[0133] It should be noted that, The term before logarithmic transformation gives the cumulative irrigation amount Adding 1 is to prevent when This ensures the stability of mathematical calculations in cases where the logarithm is negative infinity.

[0134] S3.4 One-hot coding of the "state coding" feature

[0135] For discrete "state coding" features, one-hot coding is used to convert them into binary vectors. Specifically, let the crop growth period have a total of Each category encodes the crop growth period. After one-hot encoding, a length of [length missing] is obtained. A sparse vector, wherein the value is only at the index position corresponding to the reproductive period of the current sample. The remaining positions are .

[0136] S3.5 Output standardized results

[0137] For the Each micro-region at the reference time All features are converted into standardized feature vectors with uniform dimensions and, to some extent, preserved physical correlations, denoted as... ;

[0138] Indicates the first Each micro-region at the reference time The standardized feature vector integrates all "intensity characteristics" after different physical perception standardization processes, all "cumulative" features after logarithmic scaling standardization, and "state coding" features after one-hot coding.

[0139] In a specific implementation, S4 is as follows:

[0140] When determining the appropriate irrigation amount for each leaching microzone, it is necessary to comprehensively consider the combined influence of multiple factors, including the microzone's real-time water and salt status, the drainage capacity of its associated shaft, historical cumulative effects, and the status of spatially adjacent microzones. Conventional neural networks typically treat each microzone as a completely independent entity for training, ignoring the spatial connections and system coupling between microzones due to geographical proximity and shared drainage facilities, resulting in low model training efficiency and weak generalization ability. Therefore, this invention uses a graph neural network model to abstract each leaching microzone as a node in a dynamic heterogeneous graph, and to abstract the hydraulic connection between the microzone and its associated shaft, as well as the spatial proximity between microzones, as edges in the graph. A gated physical memory unit is introduced to inject the standardized water and salt balance residuals and their historical sequences as physical memory into the node state update process. This allows the model to not only rely on the similarity of node features when transmitting information between nodes, but also incorporate the physical driving signals of the water and salt transport process, thereby achieving accurate and physically interpretable predictions of microzone irrigation demand, as shown in Figure 4. The specific steps are as follows:

[0141] S4.1 Construction of Node Physical Sensing Features

[0142] Using standardized feature vectors as the basic features of each rinsing micro-region node, to further enhance the model's ability to perceive the physical processes of water and salt transport, two additional physical perception features are extracted and constructed from the rinsing synchronization dataset and concatenated with the basic features to form the initial hidden state feature vector of each node, as follows:

[0143]

[0144] In the formula, Represents a node The hidden state feature vector at layer 0 (i.e., the input layer) is the initial encoding of that node by the graph neural network model. This represents the vector concatenation operator; Indicates the first The index of each micro-region node at time. The standardized feature vector corresponding to the reference time is defined, for the sake of simplification, as follows: That is to say, the first Each micro-region at the reference time The standardized feature vector, For the definition of equivalent symbols, Indicates the first Each micro-region at the reference time The standardized feature vector;

[0145] Indicates the first The index of each micro-region node at time. The historical water and salt pressure index corresponding to the reference time is a dimensionless scalar used to quantify the cumulative degree of water and salt stress experienced by the micro-region over a past period. A higher value indicates more severe historical water and salt stress. Its calculation depends on the standardized soil salinity and soil volumetric water content at historical times in the micro-region, and is expressed as follows: ;

[0146] This represents the graph node index, corresponding to a specific rinsing micro-region number, with a value range of [value range missing]. ; This represents the total number of nodes in the diagram, which is the total number of rinsing micro-regions within the system. This indicates the length of the historical review window, which is the number of backtracking time steps used to calculate the historical stress index. It is a preset hyperparameter, with an example value of 5. This represents the historical time offset index, used to traverse various times within the historical window; This represents the moisture content sensitivity coefficient, a preset hyperparameter used to adjust the attenuation intensity of soil moisture content in response to historical salinity pressure memory. An example value is 0.5. Indicates the first Micro-regions at historical moments Standardized soil volumetric water content; Indicates the first Micro-regions at historical moments Standardized soil salinity content. Indicates at time Depth of groundwater The determined drainage efficiency function is calculated in the same way as the drainage efficiency function. ; Indicates the index at time. The groundwater depth measurement value monitored by the shaft at the corresponding reference time; This represents the shape coefficient that influences mineralization. It is a preset hyperparameter used to adjust the sensitivity of groundwater mineralization deviation from the baseline value to the impact on drainage potential. The example value is 0.01, and its unit must ensure that the exponential term is dimensionless. Indicates the index at time. The groundwater salinity measurement value monitored by the shaft at the corresponding reference time; This represents the baseline value of groundwater salinity. It is a preset empirical parameter that represents an ideal salinity level with high salt output efficiency during drainage. An example value is 5 grams per liter.

[0147] S4.2, Physics-Guided Graph Neural Network Model

[0148] The graph neural network model is composed of A layered graph attention network is constructed by stacking layers to aggregate information about each node itself and its first-order neighbors. Layer, for the target node With any of its neighboring nodes The original attention score is calculated, which depends not only on the feature similarity between the two nodes after transformation, but also on the physical state similarity as a guiding signal, and is expressed as:

[0149]

[0150] In the formula, Indicates the first Layered network, neighbor nodes For the target node The higher the original attention score, the more important the state of the neighboring nodes is to the state of the target node during the message passing process in the current layer. This represents a linear rectified activation function with leakage, which introduces nonlinearity into the result of adding the linear transformation and the physical guide term; Indicates the first The attention vector of a layer is a trainable parameter with dimension . This is used to map the concatenated transformed features to a scalar; express transpose; Represents the target node The index of the neighboring nodes; Represents the transpose operation of a vector or matrix; The graph neural network represents the first... The layer's weight matrix, used to perform linear transformations on the node features, has a dimension of [missing information]. , are trainable parameters; Represents the hidden state feature vector The dimension; The graph neural network represents the first... The dimension of the features after the layer linear transformation; Represents a node In the graph neural network The hidden state feature vector of layer is the first layer. The output of the layer network ( ), that is, through the first The node features are updated after the neighbor information is aggregated at each layer; Represents a node In the The hidden state feature vector of layer is the first layer. The output of the layer network ( ), that is, through the first The node features are updated after the neighbor information is aggregated at each layer; Indicates will and Perform vector concatenation operation; The weight coefficient of the physics guidance term is a trainable scalar parameter used to balance the relative importance of node feature similarity and physical state similarity in the final attention allocation. The physical state similarity function is used to quantify the similarity of the water-salt balance states of two micro-region nodes at the current moment. The calculation method is expressed as follows: ; The temperature coefficient is a hyperparameter used to adjust the smoothness of similarity calculations. Represents a node Time Index The simplified water-salt balance residuals at the corresponding reference time; Represents a node Time Index The simplified water-salt balance residual at the corresponding reference time.

[0151] Use the Softmax function on the target node. The original attention coefficients of all neighbors are normalized to obtain the normalized attention weights, which are expressed as follows:

[0152]

[0153] In the formula, The graph neural network represents the first... Layer nodes For the target node The normalized attention weights are used in the graph neural network. Neighbor nodes after layer normalization For the target node Attention weights represent the proportion or contribution of information from neighboring nodes during the final aggregation. Represents the nodes in the graph The set of neighboring nodes.

[0154] Based on normalized attention weights, nodes In the The updated features of the layer are obtained by weighted aggregation of the transformed features of neighboring nodes and then nonlinear activation to obtain the updated hidden state feature vector, represented as:

[0155]

[0156] In the formula, Represents a node After the first The hidden state feature vector updated by layer encoding; This represents the exponential linear unit activation function, which introduces nonlinearity to enable the model to learn more complex patterns.

[0157] go through After layer encoding, deep encoded features containing higher-order neighbor information are obtained for each node, and defined as follows: Represents a node After the first The updated hidden state feature vector after layer encoding contains nodes. themselves and The fusion of information from all nodes within the order-neighbor range is guided by the similarity of physical states, representing the micro-region. Deep state encoding under the entire local spatial association and physical constraints, where This represents the total number of layers in the graph neural network stack, which is a preset hyperparameter.

[0158] S4.3, Dual-branch Decoding and Prediction Output

[0159] A dual-branch decoding structure is adopted for collaborative prediction and verification. The first branch predicts the irrigation demand in the future decision-making cycle, and the second branch predicts the expected changes in the core water and salt indicators after the implementation of the irrigation amount, thereby realizing the self-physical verification and closed-loop optimization of the decision.

[0160] (1) First branch

[0161] The first branch predicts the recommended irrigation amount for future decision cycles using the first multilayer perceptron, expressed as:

[0162]

[0163] In the formula, Indicates the first Micro-region at time index The predicted future irrigation demand at the corresponding baseline time, in millimeters; This refers to the first multilayer perceptron, which is a multilayer perceptron specifically designed for irrigation quantity prediction. This represents the element-wise multiplication operator; The weight vector of the first multilayer perceptron is a trainable parameter used to adapt the influence of the scalar drainage potential coefficients to each dimension of the feature vector.

[0164] It should be noted that, The prediction is based on the current reference time. Starting from this point, a future decision-making cycle (e.g.) Total irrigation demand within (days), time index Its function is to identify the system state moment on which the current decision is based; the model, according to... All synchronized and standardized data at any given time are used to make irrigation decisions for the future period.

[0165] (2) Second branch

[0166] The second branch uses a second multilayer perceptron to predict the expected change in soil salinity at future times, expressed as:

[0167]

[0168] In the formula, Indicates the predicted first The change in soil salinity in a micro-region from the current time to a set future time. This value is a standardized value. This refers to the second multilayer perceptron, which is a multilayer perceptron specifically designed for predicting the state of water and salt.

[0169] It should be noted that, The model predicts the first Each micro-region from the current reference time At a predetermined time in the future (e.g., the end of a decision-making cycle) The change in soil salinity.

[0170] In a specific implementation, as shown in Figure 5, S5 is as follows:

[0171] S5.1, Define the composite loss function

[0172] To simultaneously optimize the accuracy of irrigation quantity prediction and water-salt state prediction, and to incorporate the physical laws of water-salt balance into the form of soft constraints, the conventional mean squared error loss function only focuses on the numerical accuracy of single-objective prediction. It cannot guarantee that the irrigation quantity recommended by the model and the predicted changes in water-salt state conform to the basic physical laws. This may lead to the model outputting irrigation suggestions with small errors on the training set but which are physically infeasible or inefficient.

[0173] This invention employs a composite loss function that integrates multi-task prediction errors, physical equation residual constraints, and spatial smoothness regularization. This not only drives the model to make accurate numerical predictions but also forces the learned mapping relationship to approximately satisfy a simplified water-salt dynamics equation. Furthermore, it encourages geographically adjacent micro-regions to make similar irrigation decisions under similar conditions, thereby improving the model's physical consistency, decision rationality, and generalization robustness. The specific steps are as follows:

[0174] The physical consistency constraint loss estimates the theoretical salinity change based on the model-predicted irrigation amount, current system state, and drainage potential using simplified theoretical equations. It then penalizes the deviation between the model-predicted salinity change and this theoretical estimate, and adds a regularization term to encourage smoother irrigation decisions in adjacent micro-zones. This is expressed as:

[0175]

[0176] In the formula, This represents the physical consistency constraint loss, which aims to inject the physical laws of water-salt balance into the model training process in the form of soft constraints. This indicates the number of samples in a training batch, with each sample corresponding to a baseline time. Indicates the sample index; This represents the first estimate based on predicted irrigation volume and current conditions. The micro-region in the first The theoretical salinity change in each sample is a standardized value, estimated based on the model-predicted irrigation amount, the current water-salt balance residual, and the system's drainage potential. The core idea is that future salinity changes are primarily driven by the current water-salt imbalance and the upcoming irrigation leaching effect. The calculation method is expressed as follows: ; This represents the first proportionality coefficient, used to map the current water-salt imbalance trend to the theoretical salinity change trend. The example value is 0.02, and the unit is set to "day / mm" to ensure the consistency of dimensions in the calculation process. This represents the second proportionality coefficient, used to map unit irrigation intensity to the theoretical leaching salt reduction effect. An example value is 0.03, and the unit is set to "day / mm" to ensure the consistency of dimensions in the calculation process. Indicates the first The micro-region in the first Simplified water-salt balance residuals for each sample (corresponding to the baseline time); The statistical period length, expressed in days, indicates the cumulative amount and is used to convert total irrigation volume into average irrigation intensity. Indicates the predicted first The micro-region in the first The change in soil salinity in each sample is a standardized value; The weight coefficient representing the spatial smoothness regularization term is a hyperparameter, with an example value of 0.01. This represents the set of all edges. Specifically, each rinsing micro-region is abstracted as a node in a dynamic heterogeneous graph, and the hydraulic connection between the micro-region and its associated shaft, as well as the spatial proximity relationships between micro-regions, are abstracted as edges in the graph. This represents the total number of edges. Indicates the connection node and nodes One of the edges; Indicates the first The micro-region in the first The predicted future irrigation demand for each sample is shown in millimeters. Indicates the first The micro-region in the first The predicted future irrigation demand for each sample, in millimeters.

[0177] In practice, the actual irrigation volume label value comes from historical agricultural records or expert optimization calculations. It comes from real irrigation data in historical agricultural records (if used to train the model), or from optimized irrigation volume calculated by physical model simulation and expert rules (if used to generate supervision signals).

[0178] It should be noted that, This term represents the natural trend of salinity change in the system under the current input and output conditions, without intervention. (Input greater than output), salt has a tendency to accumulate, so this contribution is positive; otherwise, it is negative. The term characterizes the leaching effect of predicted irrigation volume on salt reduction. The greater the irrigation intensity, the stronger the salt reduction effect, thus contributing negatively. The sum of these two factors yields the predicted future irrigation demand from the current time to future times. Under the given conditions, this is the theoretical change in salinity, and the value is the value in the standardized space.

[0179] The model's total loss function is a weighted average of three parts: irrigation quantity prediction loss, water-salt state prediction loss, and physical consistency constraint loss, expressed as:

[0180]

[0181] In the formula, It represents the total loss function, which guides the updating of model parameters, so that the trained model can simultaneously meet the requirements of accuracy, physical rationality and spatial consistency. This represents the weighting coefficient of the irrigation volume prediction loss term; it is a hyperparameter with an example value of 1.0. The weighting coefficient for the water-salt state prediction loss term is a hyperparameter, with an example value of 0.8. The weighting coefficient for the physical consistency constraint loss term is a hyperparameter, with an example value of 0.5. The loss from irrigation volume prediction is represented by the Huber loss function, which is relatively insensitive to outliers, to improve the robustness of the model regression. The calculation method is expressed as follows: ; Indicates the first In the nth sample The actual irrigation volume label value for each micro-region during the future decision-making cycle, in millimeters; The loss in water and salt state prediction is represented by the mean squared error loss function, which measures the accuracy of the model's prediction of dynamic changes in water and salt. The calculation method is expressed as follows: ; This represents the Huber loss function. An example value for the threshold hyperparameter of the Huber loss function is 1.0. Indicates the first In the nth sample The actual change in soil salinity in a micro-region from the current time to a predetermined future time is calculated from the actual observations at subsequent times and converted into standardized values.

[0182] S5.2. Initiate the formal training and iterative update process of the micro-area irrigation demand prediction model, using the dataset formed after multi-source heterogeneous data cleaning and spatiotemporal synchronization, and standardization processing based on water-salt balance physical constraints as the foundation.

[0183] Before training begins, the entire dataset is first divided into training, validation, and test sets in chronological order. The training set is used to learn the model parameters, the validation set is used to monitor model performance and adjust hyperparameters during training, and the test set is used to finally evaluate the model's generalization ability.

[0184] The model parameters are initialized using standard neural network initialization methods. In each training iteration, a batch of samples is randomly selected from the training set. Each sample contains the standardized feature vectors of all leaching micro-regions at a certain baseline time and their corresponding true labels, i.e., the actual irrigation amount within the decision cycle after that time and the actual change in soil salinity observed at subsequent times. The model first performs forward propagation. Based on the input node features and the constructed graph structure, it sequentially passes through multi-layer encoding and two-branch decoding of the graph neural network to output the predicted irrigation demand and predicted salinity change for each micro-region. Then, the total loss function value of the batch of samples is calculated. The total loss is obtained by weighted summation of three parts: irrigation amount prediction loss, water-salt state prediction loss, and physical consistency constraint loss. Using the backpropagation algorithm, the gradient of the total loss with respect to all trainable parameters of the model is calculated, and an adaptive optimization algorithm is used to update the model parameters according to the gradient direction and learning rate to minimize the total loss.

[0185] The above process is repeated cyclically, and each round of traversal of the entire training set is called a training cycle.

[0186] After each training cycle, the current model is evaluated using an independent validation set. The loss value on the validation set is calculated to monitor the model's performance on unseen data and prevent overfitting.

[0187] The model will continue to be iteratively updated until a preset stopping condition is met. Stopping conditions typically include the validation set loss no longer decreasing significantly over multiple consecutive training epochs, or reaching a pre-set maximum number of training epochs. Once training stops, the model parameter version with the best performance on the validation set is saved as the final trained micro-area irrigation demand prediction model for subsequent decision-making and regulation applications.

[0188] In a specific implementation, S6 is as follows:

[0189] Once the micro-area irrigation demand prediction model is trained, it can be put into practical application to achieve intelligent decision-making and precise control of the coordinated work of drainage and saline-alkali land conversion. During actual operation, the system periodically executes the following decision-making and control cycle.

[0190] First, at each decision moment, the system automatically collects the latest information such as data from the second sensor group of each rinsing micro-zone, data from the first sensor group of the shaft, and external meteorological data. This multi-source heterogeneous data is sent to the data processing module, and the system performs cleaning, spatiotemporal synchronization and physical perception standardization according to the steps described in S5 and S6 to generate standardized feature vectors for all rinsing micro-zones at the current moment.

[0191] Then, these standardized feature vectors, combined with the constructed physical perception features and the graph structure defined according to the governance unit layout, are input into the trained micro-area irrigation demand prediction model. After forward calculation, the model outputs a recommended irrigation demand prediction value for each leaching micro-area within a future decision cycle. This prediction value not only considers the water and salt status of the micro-area itself, but also integrates the spatial correlation influence of neighboring micro-areas, the drainage potential of the shaft to which it belongs, and the historical water and salt stress memory. It is a collaborative decision-making suggestion generated under the guidance of physical laws.

[0192] After obtaining the irrigation demand forecast for each micro-zone, the control system converts it into specific execution instructions. For irrigation control, the system issues the recommended irrigation amount for each micro-zone to the corresponding rinsing solenoid valve and irrigation pump, and achieves precise and independent freshwater rinsing operations by controlling the opening duration and flow rate of the rinsing network.

[0193] Simultaneously, drainage control is implemented. The system integrates information such as the predicted total irrigation demand of all micro-areas, real-time groundwater levels, and groundwater salinity to dynamically adjust the operating strategy of the variable frequency submersible pumps in the shafts. For example, when a large-scale leaching operation is predicted, the system can increase the pumping power of the submersible pumps in advance or simultaneously to proactively lower the groundwater level, preventing the leaching water from infiltrating and causing the water level to rise above the critical depth for salt return, and accelerating the discharge of leached salts out of the treatment unit through drainage ditches. Conversely, when irrigation demand is low and the groundwater level is safe, the drainage intensity can be reduced to save energy.

[0194] Based on this, irrigation leaching and drainage / salt removal achieve spatiotemporal coordination under the unified scheduling of model-driven decision-making. After each irrigation and drainage operation, the system enters the next monitoring cycle, with sensors continuously collecting data and recording the actual irrigation volume and drainage conditions. This new data is then incorporated into the historical dataset, which can be used to periodically update and fine-tune the predictive model, enabling it to adapt to long-term changes in soil, crop, and environmental conditions. This forms a closed-loop intelligent control system of "monitoring-decision-execution-feedback-optimization," continuously and adaptively promoting the coordinated management of drainage and salt conversion in distributed coastal saline-alkali land.

[0195] Example 2

[0196] An intelligent control system for drainage and salinization improvement in coastal saline-alkali land, based on artificial intelligence, includes a zone isolation module, a subsurface drainage module, an irrigation and leaching module, a monitoring module, an intelligent collaborative control module, and a wind-solar hybrid power generation system. This is described in detail below:

[0197] The zoned isolation module includes a geomembrane laid vertically along the boundary of treatment unit 1 to block lateral saline water intrusion and create an independent environment for internal salt control. In the early stage of construction, the site to be treated was scientifically divided into several independent treatment units 1 through on-site surveys of soil salinity distribution, topography, and groundwater flow direction. Subsequently, the geomembrane was vertically laid around each treatment unit 1 to form a continuous vertical anti-seepage wall 2. The geomembrane was made of salt-alkali resistant and anti-aging material, and the laying depth reached the stable impermeable layer to ensure the anti-seepage effect. At the same time, based on the spatial variation map of soil salinity inside treatment unit 1, high, medium, and low salinity micro-zones were further refined to clearly define areas with different leaching needs, forming multiple targeted leaching micro-zones, laying the foundation for subsequent differentiated leaching operations.

[0198] The underground drainage module includes underground pipe 3, laid below the topsoil layer, for draining saline water from the soil after leaching. Underground pipe 3 is made of corrosion-resistant and pressure-resistant PVC or PE material, with evenly spaced perforations in the pipe wall and an external filter layer to prevent soil particles from clogging the pipe. It is laid in a grid or herringbone pattern, with the laying depth determined based on the local topsoil thickness and groundwater level, typically 30-80cm below the topsoil layer. This ensures efficient collection of leached saline water without affecting crop root growth or field operations.

[0199] The vertical shaft 6 drainage module includes a vertical shaft 6 and a drainage device. The drainage device includes a variable frequency submersible pump 13 and a drainage pipe 7, used to collect the brine from the underground pipe 3 and efficiently pump it to the external drainage ditch 8. The end of the underground pipe 3 is connected to the vertical shaft 6, and the variable frequency submersible pump 13 is installed in the vertical shaft 6. The submersible pump is connected to the drainage pipe 7, and the end of the drainage pipe 7 leads to the drainage ditch 8. Specifically, the ends of the underground pipes 3 are uniformly connected to the vertical shaft 6, so that the brine collected by the underground pipes 3 in each area can converge into the vertical shaft 6. Each vertical shaft 6 is equipped with a variable frequency submersible pump 13, which is connected to a controller via a cable and can automatically adjust its operating power according to the water level changes in the vertical shaft 6. The outlet of the submersible pump is connected to a high-strength drainage pipe 7, which is equipped with fixed supports and anti-backflow devices along its route, and its end leads directly to the drainage ditch 8 to ensure that the discharged brine can be quickly discharged.

[0200] The irrigation and leaching module is used to extract fresh water from the irrigation ditch 9 to leach saline-alkali land. The irrigation and leaching module includes a leaching pipe network 4, which is connected to an irrigation pump 10. A leaching solenoid valve 5 is installed on the leaching pipe network 4 for each leaching micro-zone. Specifically, the leaching pipe network 4 is reasonably laid out according to the distribution of the treatment unit 1 and the leaching micro-zone, and the pipe network branches extend to each leaching micro-zone. The water inlet of the leaching pipe network 4 is connected to the irrigation pump 10, which extracts qualified fresh water from the irrigation ditch 9. In order to achieve differentiated and precise leaching, a leaching solenoid valve 5 is installed on the leaching pipe network 4 for each leaching micro-zone, which can independently control the start and stop of leaching and the amount of leaching water according to the salinity of each micro-zone.

[0201] The monitoring module includes a first sensor group and a second sensor group, used to monitor soil moisture, water level, and water quality data, providing a basis for subsequent regulation through comprehensive data collection. The first sensor group includes a groundwater level sensor and a groundwater quality sensor, installed in shaft 6, which can monitor changes in groundwater level within shaft 6 in real time, as well as water quality indicators such as salinity and conductivity of collected saline water. The second sensor group consists of one or more of the following: soil temperature and humidity sensor, soil conductivity sensor, soil salinity sensor, and soil pH sensor. It is deployed in treatment unit 1, evenly distributed across each treatment unit 1 and different salinity micro-zones according to monitoring needs, enabling multi-point, real-time monitoring of key soil parameters and a comprehensive understanding of dynamic changes in soil moisture and salinity.

[0202] The intelligent collaborative control module includes a controller, which is connected to the drainage module of shaft 6, the irrigation and rinsing module, and the monitoring module for dynamic control of rinsing and drainage. Specifically, the controller establishes control and data transmission connections with the variable frequency submersible pump 13 of the drainage module of shaft 6, the irrigation pump 10 and rinsing solenoid valve 5 of the irrigation and rinsing module, and various sensors of the monitoring module through signal lines. During operation, the controller receives soil and water quality and water level data transmitted by the monitoring module in real time, compares and analyzes them with preset appropriate thresholds, and automatically issues control commands when the monitoring data exceeds the threshold, such as activating the rinsing solenoid valve 5 of the corresponding micro-area for rinsing, adjusting the power of the submersible pump to speed up drainage, or stopping the rinsing operation to ensure coordinated rinsing and drainage, improve salt control efficiency, and save water resources.

[0203] The wind-solar hybrid power generation system, as the core of the entire coupled system, utilizes wind and solar energy in a coordinated manner and is electrically connected to each power-consuming module for power supply. This system is electrically connected to the variable frequency submersible pump 13 of the vertical shaft 6 drainage module, the irrigation pump 10 of the irrigation and rinsing module, various sensors of the monitoring module, and the controller of the intelligent collaborative control module. Equipped with an energy storage device, the system can store excess electrical energy, effectively addressing situations with insufficient sunlight or wind, such as cloudy or rainy days or nighttime, ensuring stable operation of the entire drainage and salt control system under different meteorological conditions, while simultaneously reducing the cost and environmental impact of traditional power supply.

[0204] The wind-solar hybrid power generation system specifically includes photovoltaic panels 12 and wind turbines 11. The photovoltaic panels 12 are used to capture solar energy and convert it into electrical energy, and the wind turbines 11 are used to capture wind energy and convert it into electrical energy. The photovoltaic panels 12 and the wind turbines 11 are electrically connected to the battery 15 through a wind-solar hybrid controller 14. The wind-solar hybrid controller 14 can realize the coordinated distribution and management of the electrical energy converted from the two energy sources, and the excess electrical energy is stored in the battery 15. The battery 15 is then electrically connected to each power consumption module through an inverter 16.

[0205] Example 3

[0206] Figures 13 to 16 compare the performance of different models in predicting irrigation demand in micro-areas, comparing four machine learning models: the micro-area irrigation demand prediction model based on graph neural networks proposed in this invention (Figure 13), and three conventional models—a fully connected neural network model (Figure 14), a random forest regression model (Figure 15), and a linear regression model (Figure 15). The fully connected neural network treats each micro-area as an independent entity for training; random forest regression is an ensemble tree model; and linear regression is a simple linear model. As shown in Figures 13 to 16, the experiment uses scatter plots to illustrate the relationship between actual and predicted irrigation amounts. The horizontal and vertical axes are in millimeters. Each subplot contains two curves: the black dashed line represents the ideal prediction line, and the colored solid line represents the regression line of each model. The determination coefficient of each model is also marked in the figure. From the scatter plot distribution shown in Figure 13, it can be seen that the prediction points of the graph neural network model of this invention are most closely distributed near the ideal prediction line, the slope of the regression line is closest to one, and the determination coefficient is the highest, indicating the best prediction accuracy. The prediction points of the fully connected neural network model shown in Figure 14 are relatively scattered, with a second-highest coefficient of determination. The prediction points of the random forest regression model in Figure 15 and the linear regression model in Figure 16 show the largest deviations, with lower coefficients of determination. Experimental results show that this invention, by modeling the spatial relationships between micro-regions through graph neural networks and introducing gated physical memory units, can more accurately capture the combined effects of the micro-region's own state, the influence of neighboring micro-regions, and drainage capacity, thereby achieving accurate prediction of irrigation demand. In contrast, conventional models, by ignoring spatial relationships and physical processes, show significantly deteriorated prediction performance.

[0207] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. An intelligent regulation method for drainage and salinization improvement of coastal saline-alkali land based on artificial intelligence, characterized in that, Includes the following steps: S1. Divide the treatment units, install shafts and drainage devices in each treatment unit, and set up the first sensor group in the shafts to collect groundwater level and groundwater quality data; divide each treatment unit into multiple leaching micro-zones, and set up the second sensor group in each leaching micro-zone to collect soil moisture information. External meteorological data and agricultural records are also collected to form a dataset; S2, the multi-source heterogeneous data in the dataset are cleaned and spatiotemporally synchronized based on the monitoring time of the vertical shaft to obtain the cleaned and synchronized feature values, forming a cleaned and synchronized dataset; S3, for the data in the cleaned and synchronized dataset, a segmented physical perception standardization method is adopted to classify the features according to their physical meaning. Adaptive standardization transformations are applied to different types of features, and water-salt balance residual terms are introduced to modulate and transform the intensity characteristic features obtained from the classification. The standardized features are integrated to obtain a standardized feature vector; S4, a micro-area irrigation demand prediction model based on graph neural network is constructed. The standardized feature vector is input, and the standardized feature vector is used as a node to construct physical perception features. The initial hidden state features containing physical perception features are input into the physical-guided graph neural network model for encoding, and the deep encoded features are output. After two-branch decoding, the predicted value of recommended irrigation amount and the change in soil salinity are output; S5. Calculate the total loss function of the model and train the model based on the total loss function; S6. The trained model performs periodic decision-making and control cycles, converting the multi-source heterogeneous data collected at each decision moment into instructions after passing through the model and simultaneously performing drainage control.

2. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence as described in claim 1, characterized in that, S3 is as follows: S3.

1. Divide the synchronous washing dataset into intensity characteristic features, cumulative quantity features, and state coding features according to the physical meaning and data properties of the features; S3.

2. Perform physical perception standardization on the intensity characteristic features, using the residual term based on the simplified water-salt balance equation to modulate the traditional Z-score standardization process to obtain the standardized intensity characteristic features; S3.

3. Perform logarithmic scaling standardization on the cumulative quantity features, first performing logarithmic scaling to alleviate data skew, and then performing Z-score standardization to obtain the standardized cumulative quantity features; S3.

4. Use one-hot encoding to convert the state coding features into binary vectors; S3.

5. Integrate the standardized intensity characteristic features, cumulative quantity features, and binary vectors to obtain the standardized feature vectors of each rinsing micro-region at each reference time.

3. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence as described in claim 1, characterized in that, The specific steps for constructing physical sensing features are as follows: using the standardized feature vector as the basic feature of each rinsing micro-area node, extracting and constructing two additional physical sensing features from the cleaning synchronization dataset and concatenating them with the basic features to form the initial hidden state feature vector of each node.

4. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence as described in claim 1, characterized in that, The specific operations of a graph neural network model are as follows: A graph neural network model includes... A stacked graph attention network layer is used to aggregate information about each node itself and its first-order neighbors. Layer, for the target node With any of its neighboring nodes The original attention score is calculated by comparing the physical state with the introduced guiding signal. Based on normalized attention weights, nodes In the The updated features of the layer are obtained by weighted aggregation of the transformed features of neighboring nodes and then undergoing nonlinear activation to obtain the updated hidden state feature vector; after... After layer encoding, we obtain deep encoded features for each node that contain information about its higher-order neighbors.

5. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence according to claim 1, characterized in that, The specific operation of the dual-branch decoding is as follows: The dual-branch decoding structure is used to collaboratively predict and verify the deep coding features. The first branch predicts the recommended irrigation amount in the future decision-making cycle through the first multilayer perceptron, and the second branch predicts the expected change in soil salinity at future times through the second multilayer perceptron.

6. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence according to claim 1, characterized in that, S5 is as follows: S5.1 First, based on the prediction results of the micro-area irrigation demand prediction model and the real labels, calculate the irrigation quantity prediction loss term and the water and salt state prediction loss term. The real labels are the actual irrigation quantity and the actual water and salt state. Then, based on the model prediction results, calculate the physical consistency constraint loss term, and introduce the spatial smoothing regularization term and the theoretical residual equation. The spatial smoothing regularization term is a regularization term that encourages smooth irrigation decisions in adjacent micro-areas. The difference between the predicted salinity change and the theoretical salinity change is calculated through the theoretical residual equation. Finally, the three loss terms are weighted and added together to obtain the total loss function. S5.2 Iteratively train the model. Divide the dataset into training set, validation set and test set according to time. Initialize the model parameters with a standard neural network. Randomly sample batches of samples in each round of training. Output the prediction results through forward propagation. After calculating the total batch loss, calculate the gradient through backpropagation. Update the parameters with an adaptive optimization algorithm to minimize the loss. One iteration of the training set is one cycle. At the end of each cycle, evaluate with the validation set. Set a stopping condition. After training stops, save the optimal parameter model of the validation set, and then evaluate its generalization ability with the test set.

7. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence according to claim 1, characterized in that, S6 is as follows: The trained model periodically executes the following decision-making and control loop: First, at each decision moment, multi-source heterogeneous data is automatically collected, and then the data is cleaned, spatiotemporally synchronized, and physically standardized to generate standardized feature vectors for all rinsing micro-zones at the current moment; then, the standardized feature vectors, combined with the constructed physically perceived features and the graph structure defined according to the governance unit layout, are input into the trained micro-zone irrigation demand prediction model, and each rinsing micro-zone outputs a recommended irrigation demand prediction value for the future decision cycle; based on the irrigation demand prediction value of each micro-zone, specific execution instructions are converted, a recommended irrigation amount is issued for each micro-zone, and drainage control is performed simultaneously.

8. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence according to claim 1, characterized in that, S2 is as follows: S2.1, using the stable and low-frequency monitoring time of the sensor group installed in the shaft as the time alignment benchmark, through parameters... This represents the number of valid data acquisition moments of the sensor array in the shaft within the observation period, forming the reference time axis. ， ;in, Represents the first on the reference time axis Each data collection moment, S2.

2. For any reference time, synchronize the sensor data of each leaching micro-zone to which it belongs. Specifically, for the soil moisture information collected by the micro-zone sensors, interpolation calculation is performed using an adaptive time window weighted average based on the soil temperature change rate to obtain the synchronized soil attribute values ​​of each leaching micro-zone at each collection time on the reference time axis. S2.

3. External meteorological data are macro-meteorological data from external data sources, including raw air temperature and raw recent cumulative evaporation, which are aligned using the nearest neighbor principle. Agricultural record data are discrete records or periodic statistical values, including crop growth period codes, raw recent cumulative irrigation volume, and raw cumulative working time recorded by the vertical well submersible pump. Based on the time attributes of the agricultural record data, it is directly associated with the specific date or statistical period of each reference time, and the corresponding values ​​are extracted and aligned. S2.

4. For each reference time and each leaching micro-zone in the vertical well monitoring, generate a multi-dimensional data vector containing all aligned features, and integrate the multi-dimensional data vectors of all times and all micro-zones to form a leaching synchronization dataset.

9. The intelligent regulation method for drainage and salinization of coastal saline-alkali land based on artificial intelligence according to claim 1, characterized in that, S1 is as follows: The plot is divided into several independent treatment units; the second sensor group installed in the leaching micro-zone includes soil temperature and humidity sensors, soil conductivity sensors, soil salinity sensors and soil pH sensors; the leaching pipeline network is equipped with leaching solenoid valves in each leaching micro-zone to control leaching individually; the dataset includes soil moisture information, groundwater level data, groundwater quality data, external meteorological data and agricultural record data.

10. An intelligent regulation system for drainage and salinization improvement of coastal saline-alkali land based on artificial intelligence, characterized in that, The method for intelligent regulation and control of drainage and salinization of coastal saline-alkali land based on artificial intelligence as described in any one of claims 1 to 9 comprises: a zone isolation module including an impermeable membrane laid vertically along the boundary of the treatment unit to block lateral saline water intrusion; a submerged pipe drainage module including a submerged pipe laid below the cultivated layer to discharge saline water leached from the soil; and a vertical well drainage module including a vertical well equipped with a drainage device comprising a variable frequency submersible pump and a drainage pipe for collecting the leached saline water discharged from the submersible pipe and efficiently pumping it to an external drainage ditch; the end of the submersible pipe is connected to the vertical well, and a variable frequency submersible pump is installed in the vertical well, the submersible pump being connected to... The system includes: a drainage pipe leading to a drainage ditch; an irrigation and leaching module for drawing fresh water from irrigation ditches to leach saline-alkali land; a leaching pipeline network connected to an irrigation pump, with a leaching solenoid valve installed in each leaching micro-zone; a monitoring module including a first and second sensor group for monitoring soil moisture, water level, and water quality; an intelligent collaborative control module including a controller connected to the vertical well drainage module, irrigation and leaching module, and monitoring module for dynamic control of leaching and drainage; and a wind-solar hybrid power generation system electrically connected to each power-consuming module for supplying power.

Images