Cold region crop irrigation system based on multi-source sensing prediction

Abstract

The invention provides a cold region crop irrigation system based on multi-source sensing prediction, and belongs to the technical field of intelligent irrigation. The system comprises the steps that soil temperature, ice content, moisture and pipeline working condition data are collected through a multi-source sensor, and a physical neural network fused with ice resistance effect factors is utilized; the effective water increment of the root system layer is predicted by coupling a heat-water migration rule; the irrigation water amount, the irrigation time and the water temperature are dynamically determined in combination with a multi-objective optimization model, and the pipeline head loss is counteracted in cooperation with a flow dynamic compensation module. The system can adapt to the low-temperature environment of the cold region, accurate matching of irrigation supply and demand is achieved, the freezing and blocking risk and water resource waste are effectively reduced, and the irrigation efficiency and crop stress resistance are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent agricultural irrigation technology, in particular to a crop irrigation system in cold regions based on multi-source sensing prediction. BACKGROUND

[0002] Agricultural irrigation in cold regions (such as high latitude and high altitude regions) faces the combined challenges of freeze-thaw cycles in the non-growing season and low temperatures in the growing season. Traditional irrigation decisions are mostly based on soil moisture monitoring and meteorological water demand calculation, and the core is to predict how much water the crops need. However, in cold regions, a more fundamental physical bottleneck is that ice in the soil can significantly hinder the flow of water, i.e. ice blockage effect. This makes it possible that even if the water demand is accurately predicted, the irrigation water may fail to effectively migrate to the root layer, resulting in irrigation failure, and even cause frost damage due to surface water freezing.

[0003] The prior art predicts irrigation amount by fusing soil humidity, meteorological, and remote sensing data, but does not involve the change in water transport physical properties when soil temperature is below freezing point, and the existing decision variables are usually only irrigation amount or irrigation timing, lacking the coordinated optimization of irrigation water temperature, a key variable that can actively alleviate the ice blockage effect. In view of the above various factors, the prediction probability is low, which is not conducive to crop irrigation.

[0004] Therefore, the present application proposes a crop irrigation system in cold regions based on multi-source sensing prediction. SUMMARY

[0005] The present application provides a crop irrigation system in cold regions based on multi-source sensing prediction to solve the technical problems mentioned above.

[0006] The present application provides a crop irrigation system in cold regions based on multi-source sensing prediction, comprising: A multi-source acquisition module is used to acquire soil temperature profile, soil volume water content, soil dielectric properties, surface snow and ice coverage data, and near-surface meteorological data in real time based on a multi-source sensing network deployed in the farmland in cold regions, and simultaneously access regional weather forecast data and crop canopy multispectral image data obtained by a UAV; A water increment prediction module is used to input the multi-source data into a pre-trained physical neural network model after spatio-temporal alignment and fusion, to predict the spatio-temporal evolution process of soil stratified temperature, liquid water content, and ice content in a future set period, and simultaneously predict the effective water increment of the root layer based on the soil unsaturated hydraulic conductivity dynamically corrected by the ice blockage effect factor, wherein the physical neural network model is trained by embedding the heat conduction equation and the equation modified by the ice blockage effect factor as constraints into the corresponding loss function; a scheme generation module configured to construct a digital twin updated synchronously with the physical farmland in the virtual space based on the prediction result, wherein processes of water infiltration, redistribution and coupling with temperature field under different irrigation schemes are simulated in the digital twin, effective water increment of the root layer and freezing and runoff risk corresponding to each scheme are quantitatively evaluated, and a collaborative optimization scheme including irrigation time, irrigation water quantity and irrigation water temperature is generated by taking maximizing the effective water increment, minimizing the risk and cost as the goal; an irrigation and optimization module configured to control the irrigation system to perform variable irrigation based on the collaborative optimization scheme, and optimize the physical neural network model and the digital twin in combination with irrigation differences.

[0007] Preferably, the method further comprises: a factor determination module configured to determine an ice resistance effect factor: wherein, is the predicted ice resistance effect factor at time t and soil depth z; is the predicted ice content at time t and soil depth z; is the soil saturated water content; is the soil temperature at time t and soil depth z; is the soil water freezing temperature; is an empirical parameter based on soil texture calibration; is a reference temperature, taking 1℃ as the value; a hydraulic conductivity correction module configured to correct the unsaturated hydraulic conductivity of the soil: wherein, is the saturated hydraulic conductivity; is the effective saturation; is an empirical parameter; is the corrected hydraulic conductivity.

[0008] Preferably, the loss function of the physical neural network model is: wherein, is the error between the network prediction value and the sensor observation value; is a physical regularization term, which forces the network output to approximately satisfy the heat conduction equation with a coupled phase change latent heat source term and the equation corrected by the ice resistance effect factor; is an ice resistance consistency constraint term; is a weighting coefficient.

[0009] Preferably, the multi-source sensor network is a network system deployed at different positions and depths in the farmland in the cold region and integrating multiple types of sensors.

[0010] The preferred objective function is one that maximizes the effective water increment while minimizing risk and cost. for: ,in, The decision variable vector, and For irrigation water volume, For irrigation start time, For irrigation water temperature; This represents the predicted effective water increment in the root zone. This is a quantified value for the risk of icing or runoff. The function is the irrigation cost function; These are the weighting coefficients for each objective; This represents the maximum effective water increment in the root zone; To maximize irrigation costs; The constraints that the optimization must satisfy include: ,in, This refers to the depth range of the root system layer; The ground temperature threshold for safe irrigation; This represents the lower limit of crop demand. These are the lower and upper limits of irrigation water temperature; These represent the lower and upper limits of irrigation water volume.

[0011] Preferably, the water increment prediction module includes: The first computing unit is used for... Key parameters corresponding to soil type S, and calculation of the effective water conversion coefficient under ice resistance conditions. ; ; ; ; in, As a physical blocking factor, it characterizes the nonlinear blocking effect of relative ice content on the continuity of liquid water; The critical ice content threshold associated with soil type S. The blocking morphology index is related to soil pore morphology. The water potential-hydraulic conductivity coupling factor characterizes the effect of the synergistic effect of soil water potential and hydraulic conductivity on the efficiency of water transport to roots in the presence of ice resistance. Ice resistance effect factor Corrected soil unsaturated hydraulic conductivity, The saturated hydraulic conductivity corresponding to soil type S. This represents the predicted soil water potential at time t and soil depth z. Water level to induce crop wilting. The transport coupling coefficient related to soil texture; The second calculation unit is used to calculate the liquid water content at predicted time t and soil depth z. and Predict the change in the effective reachable water volume of the root zone from the current time t0 to the future time t. : ,in, The upper boundary of the root system, usually the surface. , This is the lower limit depth of effective water absorption by the root system. This refers to the root water absorption rate per unit time and unit volume of soil.

[0012] Preferably, the irrigation and optimization module includes: The virtual execution unit is used to simulate the collaborative optimization scheme and perform virtual pre-execution of the collaborative optimization scheme in combination with real-time acquired environmental disturbance prediction data, generating high-fidelity execution instructions containing dynamic compensation parameters. The real-time acquisition unit is used to acquire real-time data streams during the execution of the high-fidelity execution instructions and construct feature fingerprint vectors. ,in, These are the real-time deviations between the actual flow rate and actual water temperature and the simulated values, respectively. Each for all ,all Historical statistical standard deviation; It is the minimum value; This represents the rate of change of pipeline pressure. Average pressure; This is the transient infiltration resistance coefficient of the surface layer; This is a real-time evaporation correction factor; The indicator evaluation unit is used to evaluate the feature fingerprint vector. The irrigation system state is represented by a control action selected from a predefined action space, and a disturbance event is randomly matched to obtain several possible state change paths of the irrigation system in the short time domain in the future, and the execution robustness index of each path is evaluated. The directional optimization unit is used to select the path with the largest expected value corresponding to the execution robustness index, and convert the control action corresponding to the path with the largest expected value into a real-time correction instruction, and combine it with the feature fingerprint vector. Directional gradient optimization is performed on the coefficients of the physical constraint terms in the physical neural network model and the system simulation parameters of the digital twin.

[0013] Preferably, the dynamic compensation parameters are used to offset the expected deviations caused by system inertia, transmission delay, and environmental disturbances.

[0014] Compared with the prior art, the beneficial effects of this application are as follows: In the field of irrigation, the ice resistance effect factor was clearly proposed and quantified, establishing a bridge connecting permafrost physics and agricultural irrigation applications, and solving the problem that existing technologies neglect the special physical mechanisms of water transport in cold regions.

[0015] By combining physical neural networks with digital twin simulation optimization, the physical neural network ensures that the prediction results conform to physical laws, and is more robust, especially in freeze-thaw conditions where data is scarce; the digital twin simulation enables a forward-looking, multi-dimensional assessment of the consequences of irrigation decisions, and elevates decision-making from a passive response to proactive planning.

[0016] Introducing irrigation water temperature as an active decision variable into the optimization system, and co-optimizing it with irrigation timing and water volume, provides an active control method to alleviate the ice resistance effect, breaking through the limitations of traditional single decision variables and improving the reliability of predicting irrigation-related instructions for crops in cold regions.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of a cold-region crop irrigation system based on multi-source sensing prediction, as described in an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] This invention provides a cold-region crop irrigation system based on multi-source sensing prediction, such as... Figure 1 As shown, it includes: The multi-source acquisition module is used to acquire soil temperature profile, soil volumetric water content, soil dielectric properties, surface snow and ice cover data and near-ground meteorological data in real time based on a multi-source sensor network deployed in cold farmland. At the same time, it can also access regional meteorological forecast data and crop canopy multispectral image data acquired by UAVs. The water increment prediction module is used to spatiotemporally align and fuse multi-source data and input it into a pre-trained physical neural network model to predict the spatiotemporal evolution of soil stratification temperature, liquid water content, and ice content within a future set time period. Based on the soil unsaturated hydraulic conductivity dynamically corrected by the ice resistance effect factor, it simultaneously predicts the effective water increment in the root zone. The physical neural network model is trained by embedding the heat conduction equation and the equation corrected by the ice resistance effect factor as constraints into the corresponding loss function. The scheme generation module is used to construct a digital twin in the virtual space that is synchronously updated with the physical farmland based on the prediction results. In the digital twin, the process of water infiltration, redistribution and coupling with the temperature field under different irrigation schemes is simulated. The effective water increment in the root zone and the risks of freezing and runoff corresponding to each scheme are quantitatively evaluated. With the goal of maximizing the effective water increment and minimizing the risks and costs, a collaborative optimization scheme including irrigation timing, irrigation water volume and irrigation water temperature is generated. The irrigation and optimization module is used to control the irrigation system to perform variable irrigation based on the collaborative optimization scheme, and to optimize the physical neural network model and digital twin in combination with irrigation differences.

[0023] Multi-source sensor networks are network systems that integrate various types of sensors, deployed at different locations and depths in cold-region farmland. They are used to comprehensively collect multi-dimensional parameters of soil, surface, and near-surface environment. This network includes distributed temperature sensing optical fibers, arrayed temperature sensors, time domain reflectometers (TDR), frequency domain reflectometers (FDR), snow depth sensors, wind speed sensors, temperature and humidity sensors, etc. For example, in a 10-acre cold-region cornfield, sensor nodes are deployed at 5m x 5m intervals. Each node is vertically equipped with sensors at depths of 0-5cm, 5-10cm, 10-20cm, 20-40cm, and 40-60cm to acquire soil parameters at different depths. The soil sensors are deployed to the set depth using a trenching and backfilling method. The sensors and data acquisition devices are connected via wired (RS485 bus) or wireless (LoRa, NB-IoT) methods. The snow depth sensor is ultrasonic and installed on a bracket 1.5 meters above the ground. The near-surface meteorological sensor is integrated into a weather station and installed at an unobstructed location on the edge of the farmland.

[0024] A soil temperature profile refers to the continuous distribution data of soil temperature at different depths from the surface to a certain depth, reflecting the vertical variation characteristics of soil temperature. For example, the collected soil temperature profile data for 0-60cm depth are: 1.2℃ at 0-5cm depth, 2.5℃ at 5-10cm depth, etc.

[0025] Soil volumetric water content refers to the volume percentage of liquid water in a unit volume of soil, and is measured by a time domain reflectometer (TDR) or frequency domain reflectometer (FDR).

[0026] Soil dielectric properties refer to the polarization response characteristics of soil to electromagnetic waves, mainly including parameters such as dielectric constant and dielectric loss. Their values ​​are closely related to the form of water in the soil (liquid water, ice). The dielectric constant of ice (about 3-4) is much lower than that of liquid water (about 80). It is measured by a combination of time domain reflectometer (TDR) and frequency domain reflectometer (FDR).

[0027] Snow and ice cover data refers to data such as snow thickness, snow density, and ice thickness on farmland surfaces, which are used to determine the state of surface freezing and its impact on irrigation water infiltration.

[0028] Near-ground meteorological data refers to meteorological parameters within a short distance above farmland, including air temperature, relative humidity, wind speed, wind direction, solar radiation, and precipitation, which directly affect soil moisture evaporation and temperature changes.

[0029] Regional meteorological forecast data refers to meteorological forecast data released by meteorological departments covering a period of time (such as the next 7 days) in the area where farmland is located, including forecast temperature, probability of precipitation, wind speed, etc.

[0030] Multispectral image data of crop canopy acquired by drones refers to images of crop canopy in different spectral bands (such as visible light, near-infrared, and short-wave infrared) obtained by drones equipped with multispectral cameras flying over farmland. This data is used to invert crop growth status (such as leaf area index and vegetation cover) and water stress conditions. For example, if the drone is a quadcopter model, flying at an altitude of 100 meters and a speed of 5 m / s, and the multispectral camera includes four bands: 450-520nm (blue band), 520-600nm (green band), 630-690nm (red band), and 770-890nm (near-infrared band), the leaf area index obtained after processing of the captured canopy image is 3.2.

[0031] Spatiotemporal alignment refers to unifying multi-source data from different sources, at different collection times, and in different spatial locations to the same time scale and spatial coordinate system. Data fusion integrates the aligned multi-source data, eliminates redundancy and noise, and extracts effective information.

[0032] The physical neural network model is a neural network model that incorporates constraints based on physical laws. Building upon traditional neural networks, it integrates relevant physical equations as regularization terms into the loss function. This ensures that the model's output not only conforms to data patterns but also satisfies physical conservation laws, improving the model's robustness under conditions of scarce data or extreme scenarios. Specifically: Input layer: 32 dimensions (number of features after fusion of multi-source data), ReLU activation function.

[0033] Convolutional layers: 2 layers. The first layer has a kernel size of 3×3 and a number of 64 kernels. The second layer has a kernel size of 3×3 and a number of 128 kernels. Both layers use MaxPooling2D pooling.

[0034] LSTM layer: 1 layer, 256 hidden units, dropout rate 0.3.

[0035] Output layer: Fully connected layer with an output dimension of 210. The activation function is Linear. The output includes soil stratification temperature, liquid water content, and ice content for every 12 hours over the next 7 days.

[0036] During training, the heat conduction equation and the water transport equation modified by the ice resistance effect factor are embedded into the loss function to constrain the model output.

[0037] Key training parameters: Optimizer: Adam optimizer, initial learning rate 0.001, decay factor 0.95 every 50 rounds.

[0038] Number of iterations: 500 rounds, using an early stopping strategy (patience=20), that is, training stops if the validation set error does not decrease for 20 consecutive rounds.

[0039] Loss function weighting coefficients: early training phase (rounds 1-200) =0.3、 =0.2; Later stages of training (rounds 201-500) =0.5、 =0.4.

[0040] Convergence criteria: mean absolute error (MAE) of the validation set ≤ 5%, and ice content prediction error ≤ 3%.

[0041] Cross-regional validation was adopted, selecting two independent experimental fields in Mohe, Heilongjiang (high latitude) and Gonghe, Qinghai (high altitude), and collecting data for 6 months for testing, including 3 validation indicators: predicted soil stratification temperature MAE≤0.3℃, predicted liquid water content MAE≤0.02, and predicted ice content MAE≤0.01.

[0042] The future forecast period refers to a pre-defined forecast time range based on irrigation decision-making needs, typically 1-7 days into the future, adjusted according to crop growth stage and irrigation cycle. For example, setting the forecast period to 3 days into the future with a time step of 12 hours means predicting soil parameters at 6 time points within the next 3 days.

[0043] Soil stratification temperature refers to the average temperature of each layer of soil divided into several layers at a set depth. The stratification depth is consistent with the sensor deployment depth of the multi-source sensor network, facilitating comparison and verification with measured data. For example, if the soil layers are 0-5cm, 5-10cm, 10-20cm, 20-40cm, and 40-60cm, the predicted temperatures for each layer over the next 12 hours are 2.1℃, 3.0℃, 3.5℃, 4.2℃, and 5.1℃, respectively.

[0044] Liquid water content refers to the volume percentage of unfrozen liquid water in the soil. It is the form of water that crop roots can absorb and utilize. The sum of liquid water content and ice content is the total soil water content.

[0045] Ice content refers to the volume percentage of water frozen into ice in the soil. The presence of ice can block the movement of liquid water and affect crop water absorption. The spatiotemporal evolution process refers to the dynamic process of soil stratification temperature, liquid water content, and ice content changing over time and at different depths within a set prediction period. For example, over the next 3 days, the temperature in the 0-5cm layer changes from 1.2℃ to 2.1℃ to 3.0℃ to 2.5℃ to 2.0℃ to 1.8℃, while the ice content changes from 0.04 to 0.03 to 0.02 to 0.02 to 0.03 to 0.04, exhibiting diurnal fluctuations. Spatially, with increasing depth, the temperature gradually increases, while the ice content gradually decreases.

[0046] The ice resistance factor is a dimensionless parameter characterizing the degree to which ice in soil hinders the transport of liquid water. Its value ranges from 0 to 1. The closer the value is to 0, the stronger the ice resistance effect and the more difficult the transport of liquid water; the closer the value is to 1, the weaker the ice resistance effect and the smoother the transport of liquid water. This factor comprehensively considers the soil ice content and the degree to which the temperature deviates from the freezing point, and is a key parameter for correcting the unsaturated hydraulic conductivity of soil.

[0047] Soil unsaturated hydraulic conductivity refers to the ability of liquid water to conduct under the influence of water potential gradient in soil under unsaturated conditions (water content lower than saturated water content). The unit is cm / s. Its value is closely related to soil moisture content, texture, structure and ice content, and it is a core parameter for describing soil moisture transport.

[0048] The effective water increment in the root zone refers to the increase in the effective liquid water content that can be absorbed and utilized by crop roots in the crop root zone (usually 0-60cm deep, adjusted according to crop type) within a future set period. It is measured in cm and is a core indicator for judging irrigation effectiveness. Its value is the difference between the predicted effective water content in the root zone and the current effective water content.

[0049] The heat conduction equation is a physical equation that describes the process of heat transfer in soil and reflects the variation of soil temperature with time and space. The heat conduction equation used in this embodiment includes a latent heat source term for phase change to adapt to the heat changes during the ice-water phase change process in soil.

[0050] The loss function is a function that measures the deviation between the predicted value of a physical neural network model and the true value (measured data). It is used to adjust the network parameters during model training to optimize the model's prediction accuracy.

[0051] A digital twin is a digital model constructed in virtual space that fully maps to a physical farmland. This model integrates the attributes and behavioral rules of all physical entities, including soil characteristics, crop growth status, meteorological environment, and irrigation system. It can synchronize the state of the physical farmland in real time and simulate the response results under different intervention measures (such as irrigation schemes). For example, constructing a digital twin of a wheat field in a cold region could include soil stratification parameters (texture, bulk density, saturated water content, etc.), wheat growth parameters (variety, growth stage, root distribution, etc.), meteorological environment parameters (real-time temperature, wind speed, etc.), and irrigation system parameters (pipeline layout, flow valve characteristics, etc.). It can receive sensor data from the physical farmland in real time, update the virtual model's state, and simulate water transport and crop growth under different irrigation volumes, timings, and temperatures.

[0052] Synchronous updates refer to the process by which the digital twin dynamically adjusts relevant parameters in the virtual model by receiving real-time multi-source sensor data and meteorological data from the physical farmland. This ensures that the state of the virtual model is consistent with the actual state of the physical farmland, thereby guaranteeing the authenticity and reliability of the simulation results.

[0053] Simulating the infiltration, redistribution, and coupling with the temperature field of water under different irrigation schemes involves inputting different irrigation schemes into a digital twin. Through embedded water transport and heat conduction models, the simulation demonstrates the process of irrigation water infiltrating from the surface into the soil, its redistribution within the soil profile, and the interaction between water movement and the soil temperature field. Examples include the influence of irrigation water temperature on soil temperature and the effect of soil temperature on water freezing / thawing. For instance, irrigation scheme 1: irrigation at 10:00 AM, water volume 50 m³ / acre, water temperature 15℃; scheme 2: irrigation at 2:00 PM, water volume 40 m³ / acre, water temperature 12℃. The simulation in the digital twin shows that in scheme 1, the irrigation water infiltrates rapidly, reaching a depth of 30 cm within one hour. After infiltration, the water redistributes to deeper layers, resulting in uniform water content in the root zone (0-60 cm) after 24 hours. The irrigation water temperature raises the surface soil temperature by 1-2℃ without freezing.

[0054] In this embodiment, the water infiltration model adopts the Green-Ampt model, the specific expression of which is as follows: ,in, For moist frontal suction (10cm for sandy soil, 20cm for loam, 30cm for clay soil). The depth of the moistened front (cm); The infiltration intensity is expressed as cm / h.

[0055] In this embodiment, the temperature field model (including latent heat of phase change) is expressed as follows: ,in, It is the volumetric heat capacity of the soil; is the soil thermal conductivity coefficient; L is the latent heat of phase change of water; This is the density of water.

[0056] Quantitative assessment refers to the quantitative calculation and classification of the effective water increment in the root zone (i.e., the increase in available water for crops after irrigation), freezing risk (the possibility of crop frost damage due to freezing of irrigation water or soil moisture), and runoff risk (the possibility of water waste due to surface runoff caused by irrigation water not infiltrating) under each irrigation scheme. For example, the quantitative index for effective water increment is set as the ratio of actual increment to demand increment (≥1 is excellent, 0.8-1 is good, <0.8 is poor); the quantitative index for freezing risk is the duration for which the surface or root zone soil temperature is below the freezing temperature (≤1 hour is low risk, 1-3 hours is medium risk, >3 hours is high risk); the quantitative index for runoff risk is the proportion of surface runoff to irrigation water volume (≤5% is low risk, 5%-10% is medium risk, >10% is high risk). The assessment of Scheme 1 yields: an effective water increment ratio of 1.2 (excellent), a freezing risk duration of 0.5 hours (low risk), and a runoff proportion of 3% (low risk).

[0057] A collaborative optimization scheme refers to an optimal irrigation plan obtained through multi-objective optimization that comprehensively considers three decision variables: irrigation timing, irrigation water volume, and irrigation water temperature. These three variables work together to maximize the effective water increment and minimize risk and cost, rather than optimizing a single variable. For example, a collaborative optimization scheme might be: irrigation timing between 10:00-11:00 AM daily (when soil temperature rises and the risk of freezing is low), irrigation water volume of 50 m³ / acre (meeting crop water requirements without excessive runoff), and irrigation water temperature of 15℃ (alleviating ice resistance and improving water infiltration efficiency). In this scheme, the three variables work together, resulting in a 15% increase in effective water increment, a 20% reduction in risk, and a 10% reduction in cost compared to optimizing water volume or timing alone.

[0058] The irrigation system executes irrigation operations based on parameters in the collaborative optimization scheme through automatic control equipment (such as time controllers, flow valves, and heating devices), achieving precise control of irrigation timing, water volume, and water temperature. Variable irrigation means that irrigation parameters (water volume, water temperature, and timing) are dynamically adjusted according to the optimization scheme, rather than remaining fixed. For example, after receiving the collaborative optimization scheme, the irrigation system's time controller starts the irrigation pump at the set time of 10:00, and the flow valve adjusts its opening according to the set water volume of 50 m³ / acre and the irrigated area, controlling the irrigation flow rate. The irrigation continues for 5 hours; the heating device heats the irrigation water to 15℃ and monitors it in real time through a temperature sensor to ensure that the water temperature is stable within ±0.5℃.

[0059] Irrigation variance refers to the deviation between actual irrigation parameters (such as actual irrigation volume, water temperature, and start time) and the parameters set in the optimization scheme during the execution of a collaborative optimization scheme in an irrigation system, as well as the difference between actual irrigation effects (such as actual effective water increment and risk) and the effects simulated by a digital twin. For example, if the optimization scheme sets the irrigation volume to 50 m³ / mu, the actual irrigation volume is 48 m³ / mu, a deviation of 2 m³ / mu; the simulated effective water increment is 4 cm, the actual effective water increment is 3.8 cm, a deviation of 0.2 cm; and the simulated freezing risk is low, while the actual freezing risk is zero, resulting in no deviation.

[0060] Based on irrigation variance data, the weights, biases, and physical constraint coefficients of the physical neural network model, as well as the soil thermal conductivity and hydraulic conductivity calibration parameters of the digital twin simulation parameters, were adjusted to make the prediction / simulation results of the model and digital twin closer to reality, thereby improving the optimization accuracy of subsequent irrigation schemes. For example, based on irrigation variance data, it was found that the ice content predicted by the physical neural network model was 5% higher than the actual value. By adjusting the weighting coefficient of the ice resistance consistency constraint term in the model, such as from 0.2 to 0.3, and retraining the model, the ice content prediction deviation was reduced to 2%. The infiltration rate simulated by the digital twin was 10% faster than the actual rate. The surface transient infiltration resistance coefficient in the model was adjusted to make the simulated infiltration rate consistent with the actual rate.

[0061] Taking irrigation during the rice seedling raising period in the cold regions of Northeast China as an example: In early spring, the soil outside the seedling greenhouses is in a state of day-to-night freezing and thawing. Monitoring shows that the temperature of the 0-5cm soil layer varies drastically from -1℃ to 3℃ daily. Soil moisture data show significant differences near 0℃, indicating the presence of intermittent ice crystals. Drones collect multispectral images of the seedling canopy every morning to calculate the CWSI index. Regional weather forecasts provide hourly temperature, solar radiation, and precipitation probability for the next 72 hours.

[0062] After preprocessing the above data, it was input into the trained model to predict that the 0-5cm soil layer would refreeze overnight (with increased ice content) within the next 48 hours. The value will drop below 0.3, indicating extremely low hydraulic conductivity, while the temperature in the 5-15cm root layer will remain above 1℃, with zero ice content. Model output predicts that if the natural conditions continue, the effective water volume in the seedling root layer will drop to the stress threshold within the next 36 hours.

[0063] The digital twin is initialized based on the current predicted state: the decision engine pre-simulates three candidate solutions: Option A: Irrigate immediately, water temperature =8℃ (well water temperature), water volume Q=15m³ / mu.

[0064] Option B: Irrigate 12 hours later (at noon the next day). =8℃, water volume Q=15m³ / mu.

[0065] Option C: Irrigate 12 hours later. =20℃ (preheated by solar energy), Q=12m³ / mu.

[0066] The preliminary results show that Scheme A is unsuitable due to low surface temperature. In Option B, most irrigation water remains on the surface, and simulations predict an 80% probability of freezing at night, posing a high risk. Option B has a higher surface temperature, but cold water irrigation still results in some water moving slowly in the subsurface, limiting the effective water increment. The thickness was only 8mm. Option C, due to its higher water temperature, slightly increased the local soil temperature and reduced... Water infiltration is smooth. The thickness reached 10.5mm, and the risk of icing was less than 5%. After comprehensive evaluation by the optimization algorithm, scheme C was selected as the optimal formulation.

[0067] The system automatically activated its solar preheating system at noon the following day, heating the irrigation water to 20°C, and then opened the valves for precise 12mm irrigation. Data was transmitted back via a sensor network at 2, 6, and 24 hours after irrigation. Actual monitoring showed that the moisture increase in the 5-15cm soil layer matched the pre-simulation of Scheme C by 92%, but the surface cooling rate was 10% faster than the pre-simulation. The model's self-evolution module used this batch of data to fine-tune the surface albedo parameters in the twin model, thus completing one learning cycle. The system successfully avoided seedling frost damage that might have been caused by early spring irrigation, achieved effective water replenishment despite the presence of ice resistance, and improved the accuracy of subsequent simulations through closed-loop learning.

[0068] The beneficial effects of the above technical solution are as follows: The multi-source acquisition module comprehensively acquires multi-dimensional data of farmland in cold regions, providing data support for subsequent prediction and optimization; the water increment prediction module embeds a physically constrained neural network model, ensuring that the prediction of soil parameters and effective water increment conforms to physical laws, thus improving prediction accuracy; the scheme generation module obtains an optimized scheme that coordinates irrigation timing, water volume, and water temperature through digital twin simulation and multi-objective optimization, breaking through the limitations of traditional single-variable optimization; and the irrigation and optimization module achieves precise irrigation control and dynamic model optimization. Overall, it realizes precise prediction, proactive planning, and dynamic optimization of irrigation for crops in cold regions, effectively solving the problems of inefficient irrigation and high risk caused by ice resistance effects in cold regions, and improving the utilization rate of irrigation water resources and the stability of crop growth.

[0069] This invention provides a cold-region crop irrigation system based on multi-source sensing prediction, and further includes: The factor determination module is used to determine the ice resistance effect factor: ,in, The ice resistance effect factor at soil depth z at time t is the predicted value. This represents the predicted ice content at soil depth z at time t. This refers to the soil saturation water content. The predicted soil temperature at soil depth z at time t; This refers to the freezing temperature of soil water. These are empirical parameters determined based on soil quality. The reference temperature is 1℃. Hydraulic conductivity correction module, used to correct the unsaturated hydraulic conductivity of soil: ,in, The saturated hydraulic conductivity; Effective saturation; These are empirical parameters; This is the corrected hydraulic conductivity.

[0070] In this embodiment, Characterizing the degree of influence of ice content on ice resistance effect, The higher the value, the more significant the effect of ice content on ice resistance; The degree to which the temperature deviates from the freezing temperature affects the ice resistance effect. The larger the value, the greater the temperature deviation and the more significant the ice resistance effect; both are dimensionless parameters. In this embodiment, Saturated hydraulic conductivity, expressed in cm / s, is the hydraulic conductivity of soil under saturated water content. It is an inherent property of soil, related to soil texture and pore structure. Sandy soils have higher saturated hydraulic conductivity, while clay soils have lower levels. Different soil textures... The values ​​are shown in Table 1: Table 1. Different soil textures value

[0071] The beneficial effects of the above technical solution are as follows: the calculation method of the ice resistance effect factor is defined by the factor determination module, which comprehensively considers the influence of soil ice content and temperature on ice resistance, and realizes the quantitative characterization of the ice resistance effect; the hydraulic conductivity correction module dynamically corrects the unsaturated hydraulic conductivity of the soil based on the ice resistance effect factor, so that the hydraulic conductivity parameter can truly reflect the water transport capacity in the permafrost environment of cold regions, solves the problem of inaccurate hydraulic conductivity calculation caused by neglecting the ice resistance effect in existing technologies, provides reliable parameter support for the accurate prediction of the effective water increment in the root zone, and improves the scientificity and accuracy of irrigation decisions in cold regions.

[0072] This invention provides an irrigation system for cold-region crops based on multi-source sensing prediction, wherein the loss function of the physical neural network model is: ,in, This represents the error between network predictions and sensor observations. The physical regularization term forces the network output to approximately satisfy the heat conduction equation of the coupled phase change latent heat source term and the equation corrected by the ice resistance effect factor. This is an ice resistance consistency constraint term; These are the weighting coefficients.

[0073] The weighting coefficients are used to adjust the weights of the physical regularization term and the ice resistance consistency constraint term in the total loss function. Controlling the strength of constraints imposed by physical laws, The strength of the ice resistance consistency constraint is controlled by two positive numbers, ranging from 0 to 1, and adjusted based on the model training performance. For example, if the data fitting error is large in the early stages of model training, the constraint should be set... =0.3, =0.2, focusing on ensuring data fit; in the later stages of training, when the data fitting error is small, adjust... =0.5, =0.4, strengthening the consistency constraints of physical laws and ice resistance.

[0074] In this embodiment, Where N is the prediction time step; M is the number of soil layers; This represents the error weight for each parameter, with a default value of 1, which is adjusted based on data reliability. These are the observed values ​​for soil stratification temperature, liquid water content, and ice content, respectively. These are the predicted values ​​for soil stratification temperature, liquid water content, and ice content, respectively.

[0075] In this embodiment, , This is the heat conduction equation; The equation is corrected for the ice resistance effect factor; , The soil thermal conductivity coefficient; The latent heat of phase transition of water; The density of water; left side To predict the rate of change of temperature over time, the right side represents the sum of the heat conduction term and the latent heat source term of phase change. The smaller the residual, the more the predicted value conforms to the law of heat conduction.

[0076] ,in, To predict soil water potential, on the left side To predict the rate of change of liquid water content over time, the right side represents the water diffusion term. The smaller the residual, the more the predicted value conforms to the water transport pattern under ice resistance conditions.

[0077] In this embodiment, Part 1 :punish In this case, ensure that the factor is non-negative; Part Two :punish In such cases, ensure that the factor does not exceed 1; Part Three : Ensure consistency in factor calculation logic.

[0078] The beneficial effects of the above technical solution are as follows: the loss function comprehensively considers the data fitting accuracy, physical law compliance, and ice resistance consistency. By introducing a physical regularization term, the model prediction results follow the physical laws of heat conduction and water transport, avoiding the problem of violating physical laws that may occur when traditional neural networks rely solely on data fitting; the ice resistance consistency constraint term ensures the rationality of the ice resistance effect factor, further improving the robustness and prediction accuracy of the model under frozen soil conditions in cold regions, and providing reliable model support for subsequent effective water increment prediction and irrigation scheme optimization.

[0079] This invention provides a cold-region crop irrigation system based on multi-source sensing prediction, with the objective function being to maximize effective water increment and minimize risk and cost. for: ,in, The decision variable vector, and For irrigation water volume, For irrigation start time, For irrigation water temperature; This represents the predicted effective water increment in the root zone. This is a quantified value for the risk of icing or runoff. The function is the irrigation cost function; These are the weighting coefficients for each objective; This represents the maximum effective water increment in the root zone; To maximize irrigation costs; The constraints that the optimization must satisfy include: ,in, This refers to the depth range of the root system layer; The ground temperature threshold for safe irrigation; This represents the lower limit of crop demand. These are the lower and upper limits of irrigation water temperature; These represent the lower and upper limits of irrigation water volume.

[0080] objective function It is the core function of multi-objective optimization of irrigation schemes, used to measure the merits of irrigation schemes. Its objectives are to maximize effective water increment and minimize risk and cost. Weighting coefficients balance the importance of the three objectives, outputting the optimal decision variable vector. .

[0081] This refers to the total water volume for each irrigation, expressed in m³ / mu (approximately 0.16 acres). It is determined based on crop water requirements, soil moisture, and irrigation system capacity, and must be taken within the upper and lower limits of the constraints. For example, if a wheat field has a high water requirement during the jointing stage, the optimized value is Q = 50 m³ / mu, meaning 50 cubic meters of water are irrigated per mu (approximately 0.16 acres). The irrigation duration is calculated as water volume / flow rate. The PLC controller controls the opening of the flow valve and the operating time of the irrigation pump to ensure that the actual irrigation water volume meets the set value.

[0082] Irrigation start time refers to the specific time when the irrigation system is activated, usually accurate to the hour. It should be chosen during a period with higher soil temperature, lower risk of freezing, and higher crop water absorption efficiency, such as 10:00-11:00 AM. (The optimized time is...) =10:00, that is, the irrigation system is started to perform irrigation operations at 10:00 am every day.

[0083] The temperature of irrigation water, measured in °C, is a key variable for mitigating the ice resistance effect. It needs to be controlled within a suitable range to melt ice in the soil, improving water infiltration efficiency, without damaging crop roots due to excessive heat. For example, optimized... =15℃, meaning that the irrigation water is heated to 15℃ before entering the farmland.

[0084] The effects of irrigation scheme u are simulated using a digital twin, and the quantitative values ​​of freezing risk and runoff risk are calculated. A weighted average method is then used to obtain the comprehensive risk quantification value. For example, after the implementation of irrigation plan u, the quantified value of freezing risk is 0.2, and the quantified value of runoff risk is 0.1. .

[0085] Cost function The expression is set according to the actual situation, such as Where a represents the water and electricity costs per unit of water volume, b represents the heating cost per unit increase in water temperature, and c represents the fixed costs. For example, a = b= If c = 2 yuan / mu, and the irrigation scheme u = (50 m³ / mu, 10:00, 15℃), then C(u) = 10 yuan / mu.

[0086] Let be the weight coefficients for each objective, dimensionless, and ranging from 0 to 1. The weighting coefficients are used to balance the importance of the three objectives: effective water increment, risk, and cost, and are adjusted according to crop growth stage, water resource status, and economic affordability. Table 2 shows the weighting coefficients for different scenarios. Table 2 Weighting coefficients for different scenarios

[0087] Constraints are the restrictions that must be met during the optimization of irrigation schemes to ensure the feasibility and safety of the optimization schemes. These constraints include soil temperature constraints, effective water increment constraints, irrigation water volume constraints, and irrigation water temperature constraints. All constraints must be met simultaneously.

[0088] Safe irrigation ground temperature threshold Based on the crop growth experiment in cold regions, the temperature was set at 2℃ for wheat and corn, 1.5℃ for potatoes, and 1℃ for rapeseed.

[0089] Lower limit of crop demand According to the water requirement pattern of crop growth stages, the water depth is 3 cm at the jointing stage, 4 cm at the heading stage, and 3.5 cm at the grain-filling stage.

[0090] Upper and lower limits of irrigation water volume: , (Sandy soil) (Loam), 80m³ / mu (clay soil).

[0091] Irrigation water temperature upper and lower limits: =10℃, =20℃ (grain crops), 18℃ (vegetable crops).

[0092] like The range is (0, 60 cm). The predicted minimum soil temperature at each depth of the root zone is 2.5℃ > 2℃, which meets the constraint conditions; if the minimum temperature is 1.5℃ < 2℃, then the irrigation plan is not feasible and the irrigation timing needs to be adjusted (such as delaying it until the soil temperature rises).

[0093] For example, the wheat jointing stage =3cm, a certain irrigation scheme =4cm>3cm, satisfying the constraint condition; if If the diameter is 2.5cm < 3cm, then the plan is not feasible and the amount of irrigation water needs to be increased.

[0094] like , The optimized Q = 50 m³ / mu satisfies 30 ≤ 50 ≤ 60, which meets the constraint. If Q = 25 m³ / mu < 30 or Q = 65 m³ / mu > 60, then... If the water volume is not suitable, the plan is not feasible and the water volume needs to be adjusted.

[0095] like , At this point, the proposed solution is deemed unfeasible, and the parameters of the heating device need to be adjusted.

[0096] The beneficial effects of the above technical solution are: through the objective function and constraints, the optimal irrigation timing, water volume, and water temperature coordination scheme can be quickly solved, breaking through the limitation of the single variable in traditional irrigation decision-making, solving the problem of balancing effective water increment, risk, and cost in cold region irrigation, and improving the scientificity and practicality of irrigation decision-making.

[0097] This invention provides a cold-region crop irrigation system based on multi-source sensing prediction, wherein the water increment prediction module includes: The first computing unit is used for... Key parameters corresponding to soil type S, and calculation of the effective water conversion coefficient under ice resistance conditions. ; ; ; ; in, As a physical blocking factor, it characterizes the nonlinear blocking effect of relative ice content on the continuity of liquid water; The critical ice content threshold associated with soil type S. The blocking morphology index is related to soil pore morphology. The water potential-hydraulic conductivity coupling factor characterizes the effect of the synergistic effect of soil water potential and hydraulic conductivity on the efficiency of water transport to roots in the presence of ice resistance. Ice resistance effect factor Corrected soil unsaturated hydraulic conductivity, The saturated hydraulic conductivity corresponding to soil type S. This represents the predicted soil water potential at time t and soil depth z. Water level to induce crop wilting. The transport coupling coefficient related to soil texture; The second calculation unit is used to calculate the liquid water content at predicted time t and soil depth z. and Predict the change in the effective reachable water volume of the root zone from the current time t0 to the future time t. : ,in, The upper boundary of the root system, usually the surface. , This is the lower limit depth of effective water absorption by the root system. This refers to the root water absorption rate per unit time and unit volume of soil.

[0098] In this embodiment, five typical soil textures in cold regions (sandy soil, loam, clay soil, silty soil, and sandy loam) were selected. For each soil type, stratified samples were collected at a depth of 0-60cm. After removing stones and plant debris, the samples were sieved through a 2mm sieve and air-dried for later use. Ice content gradient settings: Five ice content gradients (0%, 5%, 10%, 15%, 20%) are set for each type of soil, which is achieved by controlling the total soil moisture content and freezing temperature.

[0099] Temperature gradient settings: cover typical temperature ranges in cold regions (-15℃, -10℃, -5℃, 0℃, 5℃), where 0℃ is the soil water freezing temperature T_f (default 0℃, special soils can be corrected through experiments).

[0100] Experimental groups: Each soil texture × ice content gradient × temperature gradient was divided into 1 group, and each group was repeated 3 times, for a total of 5 × 5 × 5 × 3 = 375 groups of experiments.

[0101] Soil sample preparation: Add water to the air-dried soil according to the target moisture content, stir evenly, put it into the ring cutter and let it stand for 24 hours to ensure uniform moisture distribution.

[0102] Freezing treatment: Place the ring sample into a freeze-thaw cycle test chamber and freeze it for 48 hours according to the target temperature gradient to achieve stable ice content.

[0103] Parameter determination: Soil volumetric water content was determined using a TDR300, and ice content was determined using a freeze-drying method. Saturated water content was determined using a soil moisture characteristic curve instrument. .

[0104] Data Fitting: Based on the measured data, the ice resistance effect factor was fitted using the least squares method. and , The relationship, solve Parameters; similarly, fitting is performed through hydraulic conductivity testing (vertical infiltration method). Parameters are set to ensure a goodness of fit ≥ 0.95.

[0105] This allows us to obtain typical soil empirical parameters for cold regions, as shown in Table 3: Table 3. Empirical parameters for typical soils in cold regions

[0106] Note: The parameters are applicable to high-latitude (above 45°N) and high-altitude (above 2000m) cold regions.

[0107] In this embodiment, ,in, The root distribution coefficients (0.2, 0.3, 0.25, 0.15, and 0.1 for depths of 0-60cm) are used. This represents the maximum root water absorption rate. This refers to the water holding capacity and water potential in the field, among which... The measurements were obtained using the pressure membrane method, which is an existing technology; for example, 15 bar for wheat and 12 bar for corn.

[0108] The beneficial effects of the above technical solution are: it decouples and quantifies the physical blocking of ice crystals from the attenuation of water potential-hydraulic conductivity transport efficiency, enabling the prediction model to determine the total liquid water content. In this process, the effective portion that can be absorbed by crops is dynamically identified, thereby directly calculating... This provides a precise quantitative basis for decision-making in solving the technical problem of having water in irrigation in cold regions but being unable to use it.

[0109] This invention provides an irrigation system for cold-region crops based on multi-source sensing prediction, wherein the irrigation and optimization module includes: The virtual execution unit is used to simulate the collaborative optimization scheme and perform virtual pre-execution of the collaborative optimization scheme in combination with real-time acquired environmental disturbance prediction data, generating high-fidelity execution instructions containing dynamic compensation parameters. The real-time acquisition unit is used to acquire real-time data streams during the execution of the high-fidelity execution instructions and construct feature fingerprint vectors. ,in, These are the real-time deviations between the actual flow rate and actual water temperature and the simulated values, respectively. Each for all ,all Historical statistical standard deviation; It is the minimum value; This represents the rate of change of pipeline pressure. Average pressure; This is the transient infiltration resistance coefficient of the surface layer; This is a real-time evaporation correction factor; The indicator evaluation unit is used to evaluate the feature fingerprint vector. The irrigation system state is represented by a control action selected from a predefined action space, and a disturbance event is randomly matched to obtain several possible state change paths of the irrigation system in the short time domain in the future, and the execution robustness index of each path is evaluated. The directional optimization unit is used to select the path with the largest expected value corresponding to the execution robustness index, and convert the control action corresponding to the path with the largest expected value into a real-time correction instruction, and combine it with the feature fingerprint vector. Directional gradient optimization is performed on the coefficients of the physical constraint terms in the physical neural network model and the system simulation parameters of the digital twin.

[0110] Preferably, the dynamic compensation parameters are used to offset the expected deviations caused by system inertia, transmission delay, and environmental disturbances.

[0111] Physical constraint coefficients are parameters embedded in the digital twin model to ensure that the simulation process conforms to the laws of natural physics. Their values ​​are determined based on actual physical characteristics, ensuring that the virtual simulation does not deviate from real-world logic. For example, in a digital twin of a cold-region irrigation system, the soil thermal conductivity coefficient is set to... This value corresponds to the actual thermal conductivity characteristics of local clay loam soil, avoiding discrepancies between the soil temperature change rate in the virtual model and reality. Key physical parameters such as soil thermal conductivity coefficient and pipe network head loss coefficient are determined through field experiments. Combined with typical parameter ranges in the Cold Region Agricultural Engineering Manual, statistical analysis methods are used to determine the initial coefficients of each physical constraint term. Based on the deviation feedback from long-term monitoring data, the coefficient values ​​are adjusted periodically to improve the physical consistency of the model.

[0112] System simulation parameters are a set of parameters that support the operation of a digital twin and describe the performance and environmental interaction characteristics of system components. They encompass equipment operating parameters, environmental impact parameters, etc. For example, the irrigation pump flow rate parameter in the digital twin is set as follows: The sprinkler head spray radius is set to 8 meters, and the wind speed influence coefficient in the meteorological environment is set to 0.8. These parameters together support the accurate simulation of the irrigation process in the virtual scenario. Specifically, the system collects the factory technical parameters of the irrigation equipment and the environmental impact data measured in the field, standardizes them, and enters them into the parameter library of the digital twin. Parameter association rules are established so that when a parameter changes (such as changing the sprinkler head and causing a change in the spray radius), the system automatically updates the related simulation parameters to ensure the consistency of the simulation logic.

[0113] Based on a digital twin simulation environment, irrigation parameters from a collaborative optimization scheme are input, and the simulation time step is set to 10 minutes to fully reproduce the entire process of irrigation water delivery, spraying, and infiltration. Deviation points during the simulation are captured in real time, and compensation parameters are calculated and incorporated into the execution instructions based on environmental disturbance prediction data. The high-fidelity execution instructions are irrigation operation instructions that have undergone virtual pre-execution verification and dynamic compensation correction. They possess high precision and strong anti-interference capabilities, ensuring that the actual execution effect is consistent with the optimization target. For example, the high-fidelity execution instructions for sunflower irrigation in the cold region of Inner Mongolia specify: turn on irrigation pumps No. 3 and No. 5, adjust the outlet pressure to 0.3 MPa, and set the sprinkler opening angle to 30°. Simultaneously, based on the predicted cooling trend due to environmental disturbance, a compensation instruction to increase the water temperature by 2°C is added. Specifically, the simulation results and dynamic compensation parameters are integrated, and the instruction content is formatted according to the control protocol of the irrigation equipment. Encrypted transmission is used to send the instructions to the field control terminal to ensure that no data loss or tampering occurs during instruction transmission, guaranteeing the accuracy of execution.

[0114] In this embodiment, the virtual pre-execution process is as follows: Simulation time step: 10 minutes / step, with a pre-execution duration of 1.2 times the total duration of the irrigation plan; ; ; in, This is the temperature compensation coefficient, with a value of 1.2. To predict ambient temperature; The inner diameter of the pipe (m); Pipeline head loss (m); This is a simulated ambient temperature value; The dynamic viscosity of water; This represents the length of the pipe.

[0115] Real-time data streams refer to the data sets collected in real time by sensors during the execution of an irrigation program, reflecting the system's operating status and environmental changes. This data serves as the core basis for closed-loop control. For example, during irrigation, sensors collect data every 5 minutes on pipeline pressure (e.g., 0.28 MPa), soil surface temperature (e.g., -1°C), and actual irrigation flow rate (e.g., 48...). Data such as pressure sensors, temperature sensors, and flow sensors, collected continuously, constitute a real-time data stream. Relying on devices such as pressure sensors, temperature sensors, and flow sensors in a multi-source sensor network, a unified data acquisition frequency and format are set. Various sensor data are collected through an IoT gateway, and after data preprocessing (removing outliers and filling in missing values), a standardized real-time data stream is formed and transmitted to the data processing center. The feature fingerprint vector is a high-dimensional data vector formed after feature extraction from the real-time data stream. It can condense and characterize the real-time operating status and deviations of the irrigation system. For example, in the feature fingerprint vector of a cold-region irrigation system at a certain moment, the normalized value of the flow deviation is 0.12, the pipeline pressure change rate is 0.05 MPa / h, and the surface transient infiltration resistance coefficient is 0.7. These values ​​comprehensively reflect the current irrigation flow rate being slightly lower than the set value and the soil infiltration rate being stable.

[0116] The action space is a predefined set of all possible irrigation system control operations, providing a range of selectable operations for real-time correction. For example, the action space of an irrigation system in cold regions includes 12 control operations, such as adjusting the irrigation pump outlet pressure (adjustment range 0.2-0.4MPa), switching sprinkler combinations, and changing irrigation duration (adjustment range ±10 minutes), covering key control dimensions such as flow rate, pressure, and time. Specifically, based on the equipment performance parameters of the irrigation system and the actual field control needs, all feasible control operations are identified, and the parameter range and execution conditions of each operation are clarified. These operations are then categorized and organized to form a structured action space library, facilitating the system to quickly select appropriate actions based on the actual situation.

[0117] Disturbance events refer to sudden environmental changes or system failures that may affect irrigation effectiveness during the irrigation process, covering multiple dimensions such as meteorology, equipment, and soil. For example, sudden short-term snowfall during irrigation (disturbance event 1) and a decrease in flow rate due to partial blockage of irrigation pipelines (disturbance event 2) can both affect the normal execution of the irrigation plan. Specifically, by statistically analyzing historical data on common disturbance types in cold-region irrigation, such as sudden drops in temperature, gusts of wind, and minor equipment failures, a disturbance event database is established; trigger conditions for disturbance events are set, and when a certain indicator in the real-time data stream exceeds the normal range (such as a sudden increase in wind speed to 6 m / s), the corresponding disturbance event is automatically matched.

[0118] In this embodiment, the execution robustness metrics of each path are evaluated and the expected value is determined, specifically including: For each candidate path A multi-perturbation exposure experimental field was constructed in a digital twin, and its response was evaluated by injecting three types of heterogeneous perturbations, generating a toughness spectrum matrix. : ; Among them, the matrix row index i1=1...m1 corresponds to m1 orthogonalized perturbation mode bases, which are obtained by performing eigenorthogonal decomposition on the history and synthetic perturbations. Each base represents an independent perturbation direction. Matrix column indices j1=1...n1 correspond to n1 dynamic evaluation stages, which will be executed in the time domain. The resilience was divided into n1 phases to assess how it changes over time. element The local stage toughness coefficient is calculated as follows: ,in, This is an ideal reference trajectory without disturbance. For path Under the influence of the i1th type of perturbation base The actual trajectory under the influence; This represents the theoretical limiting trajectory with the worst performance under this perturbation basis; It is the minimum value, and takes the value of ; The uncertainty in impact assessment is decomposed into random uncertainty (UA) and cognitive uncertainty (UE). UA originates from the inherent randomness of the environment and system, and is characterized by the random intensity combination of the perturbation basis. UE originates from the inaccuracy of the model (digital twin) itself, and is characterized by key physical parameters (such as the ice resistance factor). In The posterior distribution of the data is used to characterize the data, which is obtained by Bayesian updating of historical data. path Expected performance of the Brussels bar By solving a minimax optimization problem, we obtain: ,in, The vector of parameters representing the cognitive uncertainty of the model follows a distribution p1 within a confidence set P0, which is defined by a high-density region of the parameter posterior distribution. It represents a random, uncertain disturbance that follows a distribution q within a fuzzy set Q, which is defined by the magnitude boundary of the disturbance basis. In specific disturbances and model parameters Next, path The ultimate effectiveness; The calculation seeks the worst-case best value of the expected efficiency under the most unfavorable model cognitive uncertainty p1 and the most unfavorable accidental uncertainty distribution q1, thus obtaining an extremely conservative but robust expectation assessment. path Emergent comprehensive value index Instead of linear or geometric combinations of dimensions, the computation is performed using a graph neural network (GNN) model: toughness spectrum matrix Consider it as a fully connected weighted adjacency matrix, construct a path resilience graph, where nodes represent evaluation stages and edge weights represent resilience transfer characteristics between stages under specific perturbations; At the same time, the expected performance of the distributed bar will be determined. Economic costs of the path and the complexity and novelty of path control sequences. As a global feature, it is attached to the graph; A pre-trained GNN model The graph is processed by combining local and global information through message passing, and finally outputs a scalar value. : ; The GNN model is trained with a large amount of historical decision and long-term irrigation effect data. Its output essentially learns and encodes a complex, nonlinear, and potentially emergent mapping relationship between the dynamic response pattern of the path under multidimensional perturbation and the long-term real value. path Expected value of final decision It is given by the following formula: ,in, It is the Sigmoid function, used to... Normalize to the (0,1) interval; It is a path The exposed cognitive uncertainty is measured by evaluating the model parameters. When it varies within its confidence set P0 The variance of the fluctuation is used for calculation; This is the penalty coefficient.

[0119] In this embodiment, the final expected value includes not only an assessment of the path's comprehensive potential. Furthermore, it explicitly deducts the value uncertainty caused by insufficient model understanding, which makes the system more inclined to choose paths that not only perform well but are also more reliable because their performance is understood more thoroughly by existing knowledge.

[0120] In this embodiment, It quantifies the path's ability to resist specific types of disturbances at a specific stage; the closer the value is to 1, the stronger the resilience. By constructing a resilience spectrum matrix, the robustness of the path is extended from a single scalar to a dynamic response panorama of a spatiotemporal-perturbation mode, achieving fine-grained and structured evaluation. It clearly distinguishes and handles two types of uncertainty and uses minimum-maximum expectation to obtain the most robust performance estimate. This represents an advanced method for dealing with the unknown at the decision theory level. By using GNN to encode the resilience spectrum, the model can capture the complex interactions between local resilience features and their nonlinear and emergent contributions to the overall value. The risk premium caused by the model's cognitive uncertainty is explicitly subtracted from the final decision value, giving the system metacognitive capabilities.

[0121] In this embodiment, the candidate path refers to a combination of control actions that may be used in a cold-region irrigation system, such as a specific execution plan like: irrigation water volume of 45 m³ / mu + irrigation water temperature of 14℃ + start time of 11:00. The multi-disturbance exposure test field is a specialized simulation environment built in a digital twin, specifically used to test the adaptability of candidate paths under various disturbances. Heterogeneous disturbances refer to different types of disturbance factors, such as sudden drops in temperature, short-term gusts of wind, and minor blockages in irrigation pipes common in cold regions. The resilience spectrum matrix is ​​a data matrix recording the resilience performance of candidate paths under different disturbance modes and evaluation stages. Row indices correspond to independent disturbance mode bases, column indices correspond to the dynamic evaluation stage of irrigation execution, and matrix elements are local stage resilience coefficients. The disturbance mode base is an independent disturbance direction obtained by performing intrinsic orthogonal decomposition on historical and synthetic disturbance data, such as "rapid cooling disturbance base" and "short-term strong wind disturbance base," each base representing a core disturbance type. The dynamic evaluation stage divides the complete irrigation execution time into several continuous time periods, for example, dividing a 2-hour irrigation process into 4 stages in 30-minute units. The local stage resilience coefficient is a core parameter that measures the path performance at each stage under a specific disturbance. Its calculation requires the actual path trajectory under that disturbance (e.g., the soil stratification temperature change curve under a certain disturbance) and the theoretically worst-performing trajectory (e.g., the lowest possible soil temperature under extreme cooling). The minimum value is set as... This avoids situations where the denominator is zero during the calculation process.

[0122] Random uncertainty (UA) stems from the inherent randomness of the environment and system, such as random fluctuations in wind speed and minute spatial differences in soil texture during irrigation. This uncertainty cannot be completely eliminated. By combining the strengths of different perturbation model bases (e.g., setting the strength of the "cooling perturbation base" to 0.8, 1.0, and 1.2 times), the impact of this randomness on irrigation effectiveness can be simulated. Cognitive uncertainty (UE) arises from the inherent inaccuracies of the digital twin model itself, such as the ice resistance effect factor. Key parameters, such as soil ice content and temperature monitoring data from the past two years, may have discrepancies between their actual and calibrated values. Using historical irrigation data, such as soil ice content and temperature monitoring data from the last two years, a Bayesian update method is employed to obtain the posterior distribution of these key parameters (e.g., soil ice content and temperature monitoring data from the past two years). The posterior distribution of the parameters is concentrated between 0.9 and 1.1, and this distribution is used to quantitatively characterize cognitive uncertainty.

[0123] Solve a minimax optimization problem to find the path efficiency that still guarantees good results under the worst-case conditions. The confidence set P0 is the range of values ​​for the model's cognitively uncertain parameter vector, defined by the high-density region of the parameter posterior distribution, such as... The confidence set P0 is defined as the interval between 0.95 and 1.05, where the probability density is highest in the posterior distribution of the parameters. The fuzzy set Q1 represents the boundary values ​​of random uncertainty perturbations; for example, the intensity fluctuation range of the "cooling perturbation basis" is between 0.7 and 1.3 times, and this range constitutes the fuzzy set Q1. Final Performance It represents the actual irrigation effect of candidate paths under specific perturbation and model parameter combinations, such as under a 1.2x cooling perturbation basis and... With a parameter combination of 1.05, the effective water increment in the root zone reaches 3.1 cm.

[0124] The path resilience map is a network constructed by treating the resilience spectrum matrix as a fully connected weighted adjacency matrix. Nodes represent dynamic evaluation stages, and edge weights represent the resilience transfer characteristics between stages under different perturbations (e.g., the resilience coefficient changes from 0.86 to 0.91 from stage one to stage two). Global features include the expected efficiency of the distributed bar, the economic cost of the path (e.g., the total cost of water, electricity, and equipment operation for irrigation), and the complexity and novelty of the control sequence (complexity refers to the number of control actions included in the path, and novelty refers to the degree of difference between the path and the historically optimal irrigation path). The pre-trained GNN model is trained using a large amount of historical irrigation decision data (e.g., candidate paths from the past 5 years, corresponding perturbation responses, and long-term irrigation effect data). It aggregates local information and global features of the path resilience map through a message passing mechanism, ultimately outputting a scalar value as an emergent comprehensive value index. First, the resilience spectrum matrix is ​​transformed into a path resilience map. Global features such as expected efficiency and economic cost of the path are extracted and added to the map. Then, the map is input into a pre-trained GNN model. The model integrates local and global information through inter-layer message passing and outputs an emergent comprehensive value index.

[0125] The expected value is obtained by normalizing the emergent comprehensive value index and then correcting it with a cognitive uncertainty measure. The Sigmoid function is used to normalize the emergent comprehensive value index to the 0-1 interval, ensuring the comparability of indices from different paths. Cognitive uncertainty measure. This is the variance of the emergent comprehensive value index as the model parameters change within the confidence set P0 (e.g., when the parameters change between 0.95 and 1.05, the index fluctuates from 0.83 to 0.89, with a variance of 0.001). The penalty coefficient is set according to the risk tolerance of irrigation in cold regions, and is usually taken as a value between 0.1 and 0.3.

[0126] Directed gradient optimization is an optimization method that precisely adjusts the coefficients of physical constraint terms in a physical neural network model and the system simulation parameters of a digital twin based on actual operational deviations, thereby improving the adaptability of the model to the simulation.

[0127] Using the actual deviation represented by the feature fingerprint vector as the optimization target, the gradient rate of change of the deviation with respect to each parameter is calculated. According to the gradient descent direction, the coefficients of the physical constraint terms and the system simulation parameters are adjusted in small steps. After each adjustment, the optimization effect is verified by a digital twin until the deviation meets the preset requirements.

[0128] Dynamic compensation parameters are correction parameters used to offset execution deviations caused by system inertia, transmission delays, and environmental disturbances, ensuring that the actual irrigation effect closely matches the optimization target. For example, in cold-region irrigation systems, the flow rate takes 30 seconds to stabilize after the irrigation pump starts due to pipeline inertia. Setting the dynamic compensation parameter to "increase the flow rate by 10% at the initial startup" can offset the flow lag deviation caused by inertia. For predicted nighttime temperature drops, a dynamic compensation parameter of "water temperature compensation +3℃" can be set to prevent irrigation water from freezing.

[0129] The beneficial effects of the above technical solution are as follows: the fusion of digital twins and physical neural network models ensures the accuracy of prediction and simulation; dynamic compensation and directional gradient optimization mechanisms effectively offset the interference caused by special environments such as low temperatures and freeze-thaw cycles in cold regions; and the application of feature fingerprint vectors enables rapid perception and precise control of system status. In practical applications, it can not only meet the water requirements of crops at different growth stages in cold regions, but also significantly improve irrigation water utilization, reduce water waste, and reduce risks such as soil freezing and pipeline failures. It balances the dual goals of increasing yield and quality with water and energy conservation, providing reliable technical support for the sustainable development of agriculture in cold regions.

[0130] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A cold-region crop irrigation system based on multi-source sensing prediction, characterized in that, include: The multi-source acquisition module is used to acquire soil temperature profile, soil volumetric water content, soil dielectric properties, surface snow and ice cover data and near-ground meteorological data in real time based on a multi-source sensor network deployed in cold farmland. At the same time, it can also access regional meteorological forecast data and crop canopy multispectral image data acquired by UAVs. The water increment prediction module is used to spatiotemporally align and fuse multi-source data and input it into a pre-trained physical neural network model to predict the spatiotemporal evolution of soil stratification temperature, liquid water content, and ice content within a future set time period. Based on the soil unsaturated hydraulic conductivity dynamically corrected by the ice resistance effect factor, it simultaneously predicts the effective water increment in the root zone. The physical neural network model is trained by embedding the heat conduction equation and the equation corrected by the ice resistance effect factor as constraints into the corresponding loss function. The scheme generation module is used to construct a digital twin in the virtual space that is synchronously updated with the physical farmland based on the prediction results. In the digital twin, the process of water infiltration, redistribution and coupling with the temperature field under different irrigation schemes is simulated. The effective water increment in the root zone and the risks of freezing and runoff corresponding to each scheme are quantitatively evaluated. With the goal of maximizing the effective water increment and minimizing the risks and costs, a collaborative optimization scheme including irrigation timing, irrigation water volume and irrigation water temperature is generated. The irrigation and optimization module is used to control the irrigation system to perform variable irrigation based on the collaborative optimization scheme, and to optimize the physical neural network model and digital twin in combination with irrigation differences.

2. The cold-region crop irrigation system based on multi-source sensing prediction according to claim 1, characterized in that, Also includes: The factor determination module is used to determine the ice resistance effect factor: ,in, The ice resistance effect factor at soil depth z at time t is the predicted value. This represents the predicted ice content at soil depth z at time t. This refers to the soil saturation water content. The predicted soil temperature at soil depth z at time t; This refers to the freezing temperature of soil water. These are empirical parameters determined based on soil quality. The reference temperature is 1℃. Hydraulic conductivity correction module, used to correct the unsaturated hydraulic conductivity of soil: ,in, The saturated hydraulic conductivity; Effective saturation; These are empirical parameters; This is the corrected hydraulic conductivity.

3. The cold-region crop irrigation system based on multi-source sensing prediction according to claim 1, characterized in that, The loss function of the physical neural network model is: ,in, This represents the error between network predictions and sensor observations. The physical regularization term forces the network output to approximately satisfy the heat conduction equation of the coupled phase change latent heat source term and the equation corrected by the ice resistance effect factor. This is an ice resistance consistency constraint term; These are the weighting coefficients.

4. The cold-region crop irrigation system based on multi-source sensing prediction according to claim 1, characterized in that, The multi-source sensor network is a network system that integrates various types of sensors and is deployed at different locations and depths in farmland in cold regions.

5. The cold-region crop irrigation system based on multi-source sensing prediction according to claim 2, characterized in that, The objective function is to maximize the effective water increment and minimize risk and cost. for: ,in, The decision variable vector, and For irrigation water volume, For irrigation start time, For irrigation water temperature; This represents the predicted effective water increment in the root zone. This is a quantified value for the risk of icing or runoff. The function is the irrigation cost function; These are the weighting coefficients for each objective; This represents the maximum effective water increment in the root zone; To maximize irrigation costs; The constraints that the optimization must satisfy include: ,in, This refers to the depth range of the root system layer; The ground temperature threshold for safe irrigation; This represents the lower limit of crop demand. These are the lower and upper limits of irrigation water temperature; These represent the lower and upper limits of irrigation water volume.

6. The cold-region crop irrigation system based on multi-source sensing prediction according to claim 5, characterized in that, The water increment prediction module includes: The first computing unit is used for... Key parameters corresponding to soil type S, and calculation of the effective water conversion coefficient under ice resistance conditions. ; ； ； ； in, As a physical blocking factor, it characterizes the nonlinear blocking effect of relative ice content on the continuity of liquid water; The critical ice content threshold associated with soil type S. The blocking morphology index is related to soil pore morphology; The water potential-hydraulic conductivity coupling factor characterizes the effect of the synergistic effect of soil water potential and hydraulic conductivity on the efficiency of water transport to roots in the presence of ice resistance. Ice resistance effect factor Corrected soil unsaturated hydraulic conductivity, The saturated hydraulic conductivity corresponding to soil type S. This represents the predicted soil water potential at time t and soil depth z. Water level to induce crop wilting. The transport coupling coefficient related to soil texture; The second calculation unit is used to calculate the liquid water content at predicted time t and soil depth z. and Predict the change in the effective reachable water volume of the root zone from the current time t0 to the future time t. : ,in, The upper boundary of the root system, usually the surface. =0, This is the lower limit depth of effective water absorption by the root system. This refers to the root water absorption rate per unit time and unit volume of soil.

7. The cold-region crop irrigation system based on multi-source sensing prediction according to claim 1, characterized in that, The irrigation and optimization module includes: The virtual execution unit is used to simulate the collaborative optimization scheme and perform virtual pre-execution of the collaborative optimization scheme in combination with real-time acquired environmental disturbance prediction data, generating high-fidelity execution instructions containing dynamic compensation parameters. The real-time acquisition unit is used to acquire real-time data streams during the execution of the high-fidelity execution instructions and construct feature fingerprint vectors. ,in, These are the real-time deviations between the actual flow rate and actual water temperature and the simulated values, respectively. Each for all ,all Historical statistical standard deviation; It is the minimum value; This represents the rate of change of pipeline pressure. Average pressure; This is the transient infiltration resistance coefficient of the surface layer; This is a real-time evaporation correction factor; The indicator evaluation unit is used to evaluate the feature fingerprint vector. The irrigation system state is represented by a control action selected from a predefined action space, and a disturbance event is randomly matched to obtain several possible state change paths of the irrigation system in the short time domain in the future, and the execution robustness index of each path is evaluated. The directional optimization unit is used to select the path with the largest expected value corresponding to the execution robustness index, and convert the control action corresponding to the path with the largest expected value into a real-time correction instruction, and combine it with the feature fingerprint vector. Directional gradient optimization is performed on the coefficients of the physical constraint terms in the physical neural network model and the system simulation parameters of the digital twin.

8. The cold-region crop irrigation system based on multi-source sensing prediction according to claim 7, characterized in that, The dynamic compensation parameters are used to offset the expected deviations caused by system inertia, transmission delay, and environmental disturbances.

Images