Agricultural canal system intelligent water distribution system and method based on multi-modal deep learning

Abstract

The invention discloses an agricultural canal system intelligent water distribution system and method based on multi-modal deep learning, and the system comprises a multi-modal data fusion and characterization module which is used for fusing meteorological, soil, crop and canal system data, and constructing a unified spatial-temporal feature tensor through a multi-modal large model; the meteorological-water demand coupling analysis module is used for analyzing future refined net irrigation water demand by using a deep space-time diagram neural network; the physical constraint embedded channel system response modeling module is used for constructing a physical information driving model to predict channel system hydraulic response; the multi-objective collaborative optimization decision module is used for solving an optimal water distribution instruction sequence by adopting a deep reinforcement learning framework; and the instruction execution and self-adaptive feedback correction module is used for realizing instruction issuing and closed-loop regulation and control based on online learning. According to the invention, accurate conversion from meteorological data to an executable irrigation instruction is realized, the physical feasibility and engineering applicability of a water distribution scheme are ensured, and the robustness and adaptability of the system to cope with environmental changes are significantly improved.

Description

A Smart Water Distribution System and Method for Agricultural Canals Based on Multimodal Deep Learning Technical Field

[0001] This invention relates to the fields of smart agriculture and water conservancy engineering technology, specifically to an intelligent water distribution system and method for agricultural canals based on multimodal deep learning. Background Technology

[0002] The efficient use of agricultural water resources is crucial for ensuring food security and sustainable development. As the core infrastructure for water conveyance and distribution, the accuracy of canal systems in scheduling and control directly affects irrigation efficiency, crop yield, and regional water security. Currently, water allocation decisions in most irrigation districts still rely on historical experience, static irrigation regimes, or simple feedback control, making it difficult to effectively integrate and analyze high-dimensional, heterogeneous, and time-varying agricultural environmental data, such as high-resolution weather forecasts, soil moisture remote sensing, and crop growth images. This results in existing systems exhibiting delayed responses and inefficient regulation when facing sudden weather changes, spatiotemporal differences in crop water requirements, and complex canal network hydraulic constraints, easily leading to water waste or localized water shortages.

[0003] In recent years, artificial intelligence technology, especially deep learning, has been initially explored in irrigation decision-making. However, existing methods generally have the following limitations: First, the processing of multi-source data is mostly simple splicing or independent modeling, lacking the ability to mine the deep semantic relationships and spatiotemporal evolution laws between cross-modal information, and unable to establish a reliable mapping model from meteorological forecasts to refined water demand; Second, model construction often ignores the inherent physical laws and engineering constraints of canal hydraulic processes, resulting in water allocation schemes that may be theoretically optimal but are physically infeasible or difficult to implement in engineering, i.e., the so-called black box decision-making risk; Third, the system architecture is mostly open-loop, lacking online learning and adaptive correction mechanisms based on real-time operational feedback, resulting in insufficient robustness and difficulty in coping with sudden disturbances such as extreme weather or equipment malfunctions.

[0004] In summary, existing technologies have not yet achieved deep integration and collaborative intelligent decision-making across the entire chain of weather forecasting, crop water requirements, canal system response, and optimized control.

[0005] Therefore, there is an urgent need for a new technical architecture that can deeply integrate multimodal sensing information to accurately predict water demand, embed physical knowledge to ensure the feasibility and security of decision-making, and have online self-learning capabilities to adapt to dynamic environments, thereby achieving truly accurate, reliable, efficient and intelligent water distribution in agricultural irrigation systems. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent water distribution system and method for agricultural canals based on multimodal deep learning.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This application provides an intelligent water distribution system for agricultural canals based on multimodal deep learning, comprising:

[0009] The multimodal data fusion and representation module is used to acquire and fuse multi-source spatiotemporal sequence data such as weather forecasts, soil moisture, crop growth remote sensing, and canal network operation status, and extract unified feature representations with cross-modal semantic consistency through a pre-trained multimodal large model.

[0010] The meteorological-water demand coupling analysis module is coupled to the multimodal data fusion and characterization module. It is used to receive the unified feature representation and model the nonlinear coupling relationship between meteorological factors, soil moisture dynamics and crop evapotranspiration water consumption through a deep spatiotemporal graph neural network, and output refined water demand forecasts for each irrigation zone in the future multiple time periods.

[0011] The physical constraint-embedded canal system response modeling module is used to construct a deep learning proxy model with embedded physical priors based on the canal system topology, hydraulic parameters and the law of conservation of mass and momentum, so as to simulate the water level and flow propagation process inside the canal system under different gate control commands.

[0012] The multi-objective collaborative optimization decision module is connected to the water demand prediction module and the canal system dynamic simulation module, respectively. It is used to construct a sequential decision problem with water supply matching degree, operation energy efficiency and safety boundary compliance as comprehensive optimization objectives, and to solve it online using an intelligent agent or model predictive controller based on deep reinforcement learning to generate a forward-looking optimal water distribution command sequence.

[0013] The instruction execution and adaptive feedback correction module is connected to the optimization decision module. It is used to issue the optimal water allocation instruction sequence to the canal system execution mechanism and collect system status feedback and environmental observation data in real time. It uses the feedback and observation data to dynamically update the model parameters of the water demand prediction module and the canal system dynamic simulation module through an online learning algorithm.

[0014] Furthermore, the multimodal large model adopts an attention mechanism-based architecture, which includes a visual encoder for processing image data, a temporal encoder for processing temporal data, and an encoder for processing structured metadata; the features output by each encoder interact and align through a cross-modal attention mechanism to achieve deep fusion of multi-source information.

[0015] Furthermore, the graph structure of the deep spatiotemporal graph neural network model uses irrigation units as nodes, and the connection relationship between nodes is based on at least one of spatial proximity, soil property similarity, and hydrological connectivity. The model captures spatial correlations through graph convolution operations and dynamically learns the influence weights of different input features on water demand prediction through an attention mechanism.

[0016] Furthermore, the physical constraint-embedded canal system response modeling module abstracts the canal system into a directed graph and constructs a physical information neural network. During the forward propagation process, the network forces the nodes to meet the water balance through the embedded physical conservation layer and simulates the water flow resistance effect through a learnable momentum approximation layer, ensuring that the model output conforms to the basic laws of hydraulics.

[0017] Furthermore, the multi-objective collaborative optimization decision module models the water allocation process as a constrained Markov decision process; its reward function is a multi-objective weighted sum, which includes at least: a weighted water demand deviation penalty term related to the priority of irrigation units, an action smoothing penalty term to suppress frequent fluctuations in control commands, and a physical constraint violation penalty term for water level or flow exceeding the safety limit.

[0018] Furthermore, the instruction execution and adaptive feedback correction module includes a drift detection unit, which is used to continuously compare the difference between the predicted data and the actual monitoring data; when the difference is detected to continuously exceed a set threshold, an online learning process is triggered to fine-tune and update some parameters in the water demand prediction module and the canal system dynamic simulation module using recent data.

[0019] Furthermore, it also includes a credible decision interpretation and visualization module, which is used to analyze and visualize the feature attention weights in the multimodal data fusion and representation module, the feature contribution in the water demand prediction module, and the action value assessment in the optimization decision module, and generate an explanatory report describing the decision basis and the multi-objective trade-off relationship.

[0020] Secondly, this application provides an intelligent water allocation method for agricultural canal systems based on multimodal deep learning, including:

[0021] Multimodal fusion sensing steps: Integrate weather forecasts, soil moisture, crop remote sensing images, and irrigation system monitoring data; extract cross-modal features through a multimodal large model to form a unified feature representation;

[0022] Meteorological-driven water demand prediction steps: Based on the unified feature representation, the coupling relationship between meteorological, soil, and crop systems is analyzed using a deep spatiotemporal graph neural network model to predict the net irrigation water demand of each irrigation unit in the future for multiple time periods;

[0023] Physically constrained canal system simulation steps: Based on the physical structure of the canal system and the laws of hydraulics, a deep learning proxy model that integrates physical information is constructed to predict the state changes of the canal system caused by water distribution orders;

[0024] Multi-objective collaborative optimization decision-making steps: With the objectives of minimizing water demand deviation, water conveyance energy consumption, and violation of physical constraints, under the constraints of the surrogate model, the optimal water allocation command sequence is solved through deep reinforcement learning or model predictive control methods.

[0025] Closed-loop execution and adaptive learning steps: Execute the optimal water allocation command and collect real-time feedback data, and dynamically update the parameters of the water demand prediction model and the canal system simulation model through an online learning algorithm.

[0026] Furthermore, in the multi-objective collaborative optimization decision-making step, the merits of the decision are evaluated by constructing a multi-objective reward function; the reward function includes at least a deviation penalty term reflecting the degree of water demand satisfaction, a smoothing penalty term reflecting control stability, and a constraint violation penalty term reflecting the safe operation of the system.

[0027] Furthermore, in the closed-loop execution and adaptive learning steps, changes in the system or environment are identified by monitoring the continuous deviation between the predicted and actual values; when significant changes are identified, the water demand prediction model and the canal system simulation model are subjected to online incremental learning and parameter adjustment using the latest collected data sequence.

[0028] Compared with the prior art, this application has the following beneficial effects:

[0029] This invention proposes an intelligent water distribution system and method for agricultural canals based on multimodal deep learning. By configuring a modular system architecture across the entire chain, it addresses the challenges of traditional methods in deeply integrating high-dimensional heterogeneous data and the disconnect between weather forecasts and irrigation instructions. This invention combines a multimodal data fusion and representation module with a weather-water demand coupling analysis module, utilizing a pre-trained large model and a deep spatiotemporal graph neural network to achieve deep semantic association mining and spatiotemporal coupling analysis of multi-source information such as weather, soil, and crops. It establishes a reliable and interpretable mapping mechanism from accurate weather forecasts to unitized water demand. Addressing the black-box problem of existing intelligent models ignoring physical constraints, which leads to infeasible decision-making engineering, this invention uses a canal response modeling module with embedded physical constraints. It embeds hydraulic physical laws such as the Saint-Venant equations in a differentiable form into the learning model, ensuring that the generated water distribution instructions strictly adhere to the conservation of mass, conservation of momentum, and the safe operation boundaries of the canal system. This gives data-driven decision-making physical credibility and direct engineering feasibility. Finally, to address the issues of insufficient robustness caused by system response lag and lack of adaptive capabilities, this invention utilizes a closed loop consisting of a multi-objective collaborative optimization decision-making module and an instruction execution and adaptive feedback correction module. This allows for real-time solution of the multi-objective optimal solution that balances water supply, energy conservation, and safety based on reinforcement learning. Furthermore, by employing an online feedback continuous correction model, the system's dynamic adaptability and overall robustness in the face of sudden disturbances such as extreme weather, crop water demand fluctuations, and equipment malfunctions are significantly improved. Attached Figure Description

[0030] Figure 1 is a schematic diagram of the overall system architecture provided in an embodiment of the present invention.

[0031] Figure 2 is a schematic diagram of the principle framework of multimodal data fusion and representation in this invention.

[0032] Figure 3 is a schematic diagram of the workflow of the meteorological-water demand coupling analysis module in this invention.

[0033] Figure 4 is a schematic diagram illustrating the interaction between the channel system response modeling embedded with physical constraints and the multi-objective optimization decision-making in this invention.

[0034] Figure 5 is a schematic diagram of the closed-loop control framework for instruction execution and adaptive feedback correction in this invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Furthermore, in this invention, an element referred to as fixed to or disposed on another element may be directly disposed on the other element, or there may be an intermediate element. When an element is considered to be connected to another element, it may be directly connected to the other element, or there may be an intermediate element present simultaneously. The terms vertical, horizontal, left, right, and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0037] Example 1

[0038] In a large, modern irrigation district in the North China Plain, where winter wheat and summer maize are the main crops, the complex canal network includes multi-level channels and dozens of automated control gates and diversion gates. Irrigation district management faces highly spatiotemporally variable meteorological conditions, diverse crop planting structures, and stringent water-saving and efficiency-enhancing targets. Traditional water allocation methods based on experience or simple models struggle to deeply integrate high-dimensional weather forecasts, remote sensing images, and real-time monitoring data, and the decision-making process lacks physical interpretability, exhibiting insufficient adaptability in the face of extreme weather events. This embodiment applies the agricultural canal system precision water allocation system and method based on multimodal large-scale model-deep learning-integrated meteorological forecasting, as described in this invention, aiming to construct an intelligent water allocation hub that integrates data-driven approaches with physical laws.

[0039] This embodiment elaborates on the overall system deployment and modular architecture:

[0040] Referring to Figure 1, the physical architecture of this system includes a high-performance computing server cluster deployed at the irrigation district dispatch center to run the core algorithm model; a soil moisture sensor network and field weather stations are deployed in the fields, and multispectral aerial photography is conducted regularly using drones; water level gauges, flow meters, and electric gate actuators with feedback are installed along the canal system. All monitoring data is transmitted back to the dispatch center in real time via a 5G / IoT private network.

[0041] The system employs a rigorous modular design at the software level, with each module working collaboratively to form an intelligent closed loop from data perception to decision execution. Core modules include: a multimodal data fusion and representation module, a meteorological-water demand coupling analysis module, a canal system response modeling module with embedded physical constraints, a multi-objective collaborative optimization decision-making module, an instruction execution and adaptive feedback correction module, and a reliable decision interpretation and visualization module to enhance system transparency. The modules exchange data and transmit instructions efficiently via a high-speed data bus and message queue.

[0042] Furthermore, this embodiment provides a detailed description of the workflow and technical implementation of the core module:

[0043] Implementation of the Multimodal Data Fusion and Representation Module: This module serves as the data hub of the entire system, responsible for accessing, fusing, and representing multi-source heterogeneous data. Data Access: Through a standardized data interface, the module acquires gridded numerical weather forecast products with a spatial resolution of 1 km for the next 7 days from the provincial meteorological center every hour. Data fields include hourly precipitation, temperature, relative humidity, wind speed, and downward shortwave radiation. This data constitutes the meteorological forecast data required by the system, specifically high-resolution numerical weather forecast gridded data for the next 1 to 7 days, including precipitation, temperature, humidity, wind speed, and radiation. Simultaneously, the module receives time-series data on the volumetric water content of the 0-100 cm soil profile from soil moisture sensors in 500 irrigation units distributed throughout the irrigation area every 30 minutes; this is soil moisture time-series data collected by soil moisture sensors deployed in the fields. Twice a week, the module receives multispectral images covering the entire irrigation area collected by drones, which, after preprocessing, generate the normalized vegetation index for each irrigation unit. Leaf area index and canopy temperature distribution maps constitute crop growth image data, specifically multispectral images of the crop canopy taken periodically by drones or fixed cameras; real-time monitoring data of the canal system flows in continuously at a frequency of once per second, including the water level before and after all gates, the flow rate through the gates, and the gate opening feedback value, i.e., data collected by water level gauges, flow meters, and gate opening sensors along the canal system; the module also has a pre-built structured knowledge base containing information such as crop type, variety, sowing date, water requirement characteristics (crop coefficient) at each growth stage, and drought and salt tolerance thresholds, as static metadata.

[0044] Multimodal Large-Scale Model Feature Extraction and Fusion: Faced with massive data streams of varying modalities and spatiotemporal scales, this module employs a pre-trained multimodal large-scale model for deep semantic understanding and unified representation. This large-scale model adopts a Transformer-based architecture and includes three core encoder branches:

[0045] Visual encoder: Using the Vision Transformer structure, the crop multispectral image is segmented and embedded and self-attention is calculated to extract deep semantic feature vectors that characterize crop growth, stress status (such as water stress and salt stress) and biomass. Specifically, it extracts canopy coverage, leaf area index and stress features.

[0046] Time-series encoder: Employs a Transformer encoder layer with position coding to process the weather forecast time series and soil moisture history time series of each irrigation unit in parallel, capturing their periodicity, trend and abrupt change patterns, thereby extracting their periodic and trend characteristics.

[0047] Text / Numerical Encoder: Processes structured metadata such as crop type, growth stage, and soil type, and transforms it into dense vectors.

[0048] The feature vectors output by each encoder are fed into a cross-modal attention fusion layer. In this layer, the system dynamically calculates the cross-attention weights between image features, temporal features, and attribute features for each irrigation unit. For example, the system automatically learns that during the crop heading stage, the correlation between canopy temperature features and future temperature forecast temporal features should be given higher weight. This dynamic cross-modal interaction and alignment enables intelligent adjustment of the contribution of different modal information.

[0049] Specifically, for visual feature vectors Temporal feature vectors and text / numerical feature vectors , fusion to generate a unified spatiotemporal feature tensor The process can be modeled as follows:

[0050] .

[0051] in, These are the attention weight matrices for visual features to temporal features, temporal features to visual features, and text / numerical features to joint visual and temporal features, respectively. Agg represents an aggregation function that achieves nonlinear mapping and fusion of features through a multilayer perceptron.

[0052] Ultimately, the module generates a unified, high-dimensional spatiotemporal feature vector for each irrigation unit. The feature vectors of all irrigation units are arranged according to their spatial location and extended along a time dimension spanning the next 168 hours (7 days), thus constructing the core, unified spatiotemporal feature tensor of the system. The spatial dimension of this tensor precisely corresponds to the irrigation unit grid, while the time dimension covers historical observations and future prediction periods, providing a standardized data foundation for all subsequent advanced analyses.

[0053] Furthermore, the implementation of the meteorological-water demand coupled analysis module in this embodiment is described in detail:

[0054] This module receives spatiotemporal feature tensors from upstream sources, and its core task is to analyze the refined net irrigation water demand of each irrigation unit over multiple future time periods.

[0055] Construction of a deep spatiotemporal graph neural network: The core of this module is a deep spatiotemporal graph neural network.

[0056] Graph Structure Definition: First, a dynamic graph structure is constructed based on the spatial location of irrigation units, soil texture similarity, and canal hydraulic connectivity. Each irrigation unit is a node in this graph, and the connections between nodes are defined by spatial proximity, soil texture similarity, and hydrological connectivity. For example, strong connections are established between two spatially adjacent units with similar soil types.

[0057] Node feature initialization: The initial features of each graph node, that is, the feature vector extracted from the spatiotemporal feature tensor of the corresponding irrigation unit.

[0058] The network contains two core layers: spatiotemporal graph convolutional layers and graph attention layers.

[0059] Spatiotemporal graph convolutional layers simulate the spatial propagation and interaction effects of meteorological conditions (such as rainfall and evaporation) and soil moisture by aggregating information in the neighborhood of the graph. For example, heavy rainfall forecast information from an upstream irrigation unit can affect the water demand calculation of adjacent downstream units through graph convolution operations.

[0060] The graph attention layer dynamically calculates the contribution weights of each node's different input features (such as leaf area index from the visual encoder and future evaporation potential from the temporal encoder) to the final water demand prediction. This enables the model to dynamically learn the relative importance of different meteorological factors (such as rainfall and evaporation) on crop water demand.

[0061] By stacking multiple layers of spatiotemporal graph convolutions and graph attention layers, the network can model meteorological factors-driven processes, soil moisture regulation, and crop physiological responses end-to-end through a highly nonlinear dynamic coupling process, without relying on traditional, simplified empirical formulas.

[0062] Probabilistic water demand prediction output: The network's output layer is for each irrigation unit. In every period of the future Generate a probability distribution of net irrigation water demand. The output consists of two values: the mean water demand. and prediction variance Mean This serves as the optimal water demand estimate for the unit during that period, guiding water allocation; while the variance... This quantifies the uncertainty of the forecast. This uncertainty mainly stems from two aspects: the inherent uncertainty of weather forecasts themselves, and the limitations of the model's understanding of the complex interactions between weather, soil, and crops; this probabilistic output provides a key input for subsequent optimization decisions based on risk perception.

[0063] Furthermore, this embodiment describes the channel system response modeling module with embedded physical constraints:

[0064] This module aims to build a physically reliable digital canal system simulator for accurately predicting the hydraulic response of the canal system under different water distribution commands.

[0065] The module first abstracts the actual canal network as a directed graph. In the graph, nodes represent canal junctions, branching points, or gate control points, and directed edges represent canal segments connecting the nodes. Each edge is associated with its physical parameters, such as length, bottom width, slope coefficient, and Manning roughness coefficient.

[0066] Based on this graph structure, a physical information-driven deep learning model is constructed. The core innovation of this model lies in embedding the hydrophysical laws represented by the Saint-Venant equations into the network architecture in a differentiable form.

[0067] The forward propagation process of the network implicitly solves for a simplified form of the Saint-Venant equations. The input is a sequence of flow or water level commands from the upstream nodes.

[0068] The network contains two types of custom layers: a physical conservation layer and a momentum approximation layer.

[0069] The physical conservation layer is a parameter-free, hard-constraint layer designed to enforce that, under all circumstances, the total inflow to each node at any given moment equals the sum of the total outflow and the change in the node's water storage, thus strictly guaranteeing mass conservation. This is typically achieved by constructing linear equality constraints and employing a projection method in the network fronthaul.

[0070] The momentum approximation layer is a learnable network layer whose inputs are hydraulic variables such as upstream and downstream water level differences and flow rates, and whose output is the energy loss caused by simulated flow resistance and inertial effects. The output of this layer is explicitly constrained within a reasonable range derived from hydraulic principles such as the Manning formula, thereby introducing prior physical knowledge and preventing the model from making predictions that violate common sense physics.

[0071] The training of the physics-guided training strategy model uses a multi-task loss function:

[0072] .

[0073] It is a data fitting term that calculates the mean square error between the network-predicted water level and flow rate values ​​and the historical actual monitoring data. It is the physical residual term. The variables such as water level and flow rate predicted in the intermediate network are substituted into the discrete form of the Saint-Venant equations and the sum of squared residuals is calculated. It is a trade-off coefficient used to balance the accuracy of data fitting with the degree of satisfaction of physical laws.

[0074] Through this training method, the model can not only learn empirical patterns from historical data, but also ensure that its predictions strictly follow the basic physical principles such as conservation of mass and conservation of momentum, thus becoming a reliable simulator in engineering.

[0075] Furthermore, this embodiment describes the implementation of the multi-objective collaborative optimization decision-making module:

[0076] Based on accurate water demand forecasting and physically reliable canal system simulation, this module is responsible for generating the optimal water distribution command sequence.

[0077] The problem is modeled as a constrained Markov decision process: the module starts in each control cycle and models the water allocation decision as a constrained Markov decision process.

[0078] state space :Include The spatiotemporal characteristic tensor at any given time (reflecting the environment), the real-time water level and flow rate of all nodes in the canal system (reflecting the system state), and the mean and variance of future multi-period water demand forecasts output by the meteorological-water demand coupling analysis module (reflecting future demand).

[0079] Action space : A continuous vector representing the target opening adjustment of all controllable gates in the next time period.

[0080] State transition: Defined by the canal response modeling module embedded with physical constraints. Given the current state and the action to be performed, the canal state at the next moment is accurately predicted by this physical information model.

[0081] Multi-objective reward function design: reward function It is carefully designed as a multi-objective weighted sum to guide the agent to learn to balance multiple conflicting objectives: ;

[0082] The first item is the core reward for meeting water demand. It is the predicted average water demand of irrigation unit j during time period t. It maps the actual amount of water that can be delivered to the unit using a canal network model. Is it related to the priority of this unit? positive correlation coefficient Higher-priority units that fail to meet their water demand will incur greater penalties. The second penalty is for smoother operation, which directly penalizes the adjustment range of the gate opening. This is to reduce equipment wear and energy consumption. The third item is the penalty for violating physical constraints. Represents the violation of the k-th safety constraint (such as the water level at a node exceeding the warning line or the flow rate of a channel section exceeding the design capacity). The ReLU function ensures that penalties are only incurred when a violation occurs. , , Adjustable weights are used to balance water supply security, operational economy, and system safety. The module employs a proximal policy optimization algorithm to train a policy network. This policy network takes the current state as input and directly outputs the optimal action probability distribution that satisfies the action range constraints. After training, this policy network can generate the optimal water allocation command within milliseconds based on the real-time state, which satisfies multiple physical constraints and dynamically balances multiple objectives.

[0083] Alternatively, a multi-objective collaborative optimization decision-making module can be connected to the meteorological-water demand coupled analytical module and the canal system response modeling module with embedded physical constraints. This module aims to minimize water demand deviation, water conveyance energy consumption, and penalties for violating physical constraints. Under the constraints of the canal system response modeling module with embedded physical constraints, it employs deep reinforcement learning or model predictive control frameworks to solve for the optimal water allocation command sequence for each control node over multiple future time periods. The optimization problem is formulated as a constrained Markov decision process, with a state space... The action space includes the spatiotemporal feature tensor at time t, the real-time state of the canal system, and future water demand predictions. The state transition is defined by the channel system response modeling module embedded in the physical constraints, representing the opening adjustment amount of each control gate, and the reward function. Designed for multi-objective weighted sum:

[0084] ;

[0085] The first item is the weighted water demand deviation penalty. and These are the water demand and the actual water distribution, respectively. Prioritization with irrigation units The correlation coefficient satisfies The second item is a penalty for smooth control of actions, and the third item is a penalty for violating physical constraints. Represents the degree of violation of the k-th physical constraint. , , As the weight coefficients, the module trains a policy network using either proximal policy optimization or the soft actor-commentator algorithm. The network directly outputs the optimal action probability distribution that satisfies the constraints;

[0086] The instruction execution and adaptive feedback correction module is connected to the multi-objective collaborative optimization decision module. It is used to issue the optimal water allocation instruction sequence to the canal system execution mechanism and collect the execution status and external environment feedback in real time. Through an online learning mechanism, it dynamically updates the parameters of the multimodal data fusion and representation module, the meteorological-water demand coupling analysis module, and the canal system response modeling module with embedded physical constraints to achieve closed-loop adaptive control.

[0087] The module employs deep reinforcement learning algorithms, such as proximal policy optimization or the soft actor-critic algorithm, to train a policy network. The policy network is in its current state. As input, it directly outputs the optimal action probability distribution that satisfies constraints such as gate mechanical amplitude limitation; after training, the strategy network can generate the optimal water distribution command sequence within milliseconds based on the real-time collected system status, which strictly satisfies multiple physical and safety constraints, and intelligently balances multiple objectives such as water supply security, operating efficiency and system safety.

[0088] Furthermore, this embodiment describes the implementation of the instruction execution and adaptive feedback correction module:

[0089] This module is responsible for translating intelligent decisions into actual control actions and giving the system the ability to continuously learn and adapt to changes in the environment.

[0090] Command issuance and execution: The module converts the optimal opening adjustment amount generated by the decision module into a specific pulse control signal through the industrial Internet of Things protocol, and sends it to the programmable logic controller of each gate, driving the motor to precisely adjust the gate to the target position.

[0091] Real-time feedback monitoring and drift detection: While executing commands, the module activates a high-frequency data monitoring thread to collect feedback from three sources in real time: 1. Actual gate opening feedback; 2. Measured values ​​from water level gauges and flow meters along the channel; 3. Real-time meteorological data such as actual rainfall and temperature. The module also embeds a concept drift detector to continuously compare the predicted and measured values ​​of key indicators.

[0092] Compare the differences between the reference crop evapotranspiration predicted by the meteorological-water demand coupled analysis module and the value calculated based on actual meteorological data.

[0093] Compare the differences between the water level and flow rate predicted by the canal system response modeling module with embedded physical constraints and the actual monitoring values.

[0094] Online learning and adaptive updates: When the drift detector detects that the mean of the sliding window of the aforementioned differences exceeds a preset threshold for multiple consecutive periods, or when the system receives a sudden extreme weather warning issued by the meteorological department, it is determined that the operating environment has changed significantly and the original model may be inaccurate. At this time, the module immediately triggers the online learning process:

[0095] This process extracts high-frequency actual operation data sequences from the cache over a recent period (e.g., 24 hours).

[0096] Supervised fine-tuning was performed on the parameters of the final layer graph attention network of the meteorological-water demand coupled analytical module and the momentum approximation layer parameters of the canal system response modeling module with embedded physical constraints, using a very small learning rate. Simultaneously, the policy network of the multi-objective collaborative optimization decision module was also fine-tuned. The parameters will also be updated incrementally with several steps based on the latest state, action, and reward sequence.

[0097] This process is efficient and incremental, enabling the core model parameters to adapt to new environments or operating conditions within minutes to tens of minutes. This achieves closed-loop adaptive control of the system, ensuring that it maintains high-precision decision-making capabilities in the face of sudden weather changes, crop variety changes, or gradual changes in equipment performance during long-term operation.

[0098] Furthermore, this embodiment elaborates on the implementation of the credible decision interpretation and visualization module:

[0099] To enhance the credibility and acceptability of the intelligent system in actual engineering management, this system integrates a credible decision interpretation and visualization module.

[0100] Multi-source information analysis: This module receives and deeply analyzes key intermediate results throughout the decision-making chain.

[0101] The weight matrix of the cross-modal attention layer is extracted from the multimodal data fusion and representation module.

[0102] By analyzing the weights of the attention layer in the analytical graph of the meteorological-water demand coupling analytical module, the water demand influence relationship among different irrigation units is analyzed.

[0103] The multi-objective collaborative optimization decision-making module calculates the value estimate of each gate action by the policy network and decomposes the contribution of different reward items to the value, wherein the reward item is at least one of water demand satisfaction, action smoothness, and safety.

[0104] Based on the above analysis, the module automatically generates a structured and readable decision explanation report, which is then projected in a visual form onto the large screen of the irrigation district dispatch center.

[0105] Using a geographic information system as the base map, layers such as future rainfall forecasts, soil moisture distribution, and crop water and heat requirements are overlaid and displayed.

[0106] The heat map visually demonstrates which geographical regions' crop image characteristics and which meteorological elements during forecast periods had a key impact on the final water demand forecast in this decision-making process.

[0107] Clearly explain the logic behind the recommended water allocation scheme in the form of cause-and-effect diagrams or text summaries. For example: because there is a forecast of heavy rainfall in area A in the next 6 hours, the current water allocation for area A is reduced; in order to ensure the high-priority winter wheat in area B, which is in the grain filling stage, its needs are given priority within the channel flow constraints.

[0108] This deep interpretability not only greatly enhances the trust of irrigation district managers in the intelligent decision-making process, but also provides them with accurate and intuitive information support for necessary human intervention in special circumstances, such as temporarily adjusting crop priorities and setting special operating modes.

[0109] Through the implementation of the complete technical chain and closed loop consisting of multimodal fusion perception, deep coupling analysis, physical constraint modeling, reinforcement learning optimization, online adaptive correction, and reliable interpretation visualization, this system has achieved unprecedented precision, intelligence, and adaptive water allocation in the irrigation area of ​​the North China Plain. The system can automatically learn and quantify the complex coupling relationships between meteorology, soil, crops, and canal systems from massive and heterogeneous agricultural environmental data, generate optimal water allocation instructions that are both physically feasible and achieve dynamic balance among multiple objectives, and continuously self-optimize based on real-time feedback. In a complete irrigation season, the system achieved water savings of approximately 15%, reduced the water shortage index of high-priority crops during their critical growth period by 40%, successfully coped with multiple sudden heavy rainfall events, effectively avoided the risk of canal overflow, and significantly improved the intelligence, precision, and reliability of water resource management in the irrigation area.

[0110] Example 2

[0111] In a mixed irrigation area combining drip and furrow irrigation in the arid Northwest region, the main crops grown are high-value cash crops such as cotton and processing tomatoes. The water source is a combination of reservoir and groundwater. Soil salinization is a significant problem in this area, requiring not only water conservation but also strict control of root zone soil salinity, with stringent requirements for irrigation water quality and leaching regimes. Traditional methods struggle to achieve a dynamic balance between multiple conflicting objectives such as water conservation, salinity control, and ensuring crop physiological water needs. This embodiment applies the system of this invention to demonstrate its powerful scalability and adaptability in addressing more complex agricultural environmental problems.

[0112] The system hardware architecture and software module composition of this embodiment are completely consistent with those of Embodiment 1; its innovation lies in the fact that, for the special needs of water and salt regulation, the data input, model analysis objectives and optimization functions have been precisely enhanced and extended, and all extensions strictly follow the module framework and method steps defined in the claims of this invention.

[0113] This embodiment provides a detailed description of the expansion of the multimodal data fusion and representation module:

[0114] In addition to the existing data interfaces of the new data source module, two additional types of key data have been added: real-time soil salinity data in the root zone collected by the soil conductivity sensor network; and water quality online monitoring instruments installed at reservoir outlets, irrigation wells, and other locations, collecting irrigation water mineralization data.

[0115] The pre-trained multimodal large model adds a dedicated temporal encoder branch to the original three encoder branches (visual, temporal, and text / numerical) to process soil salinity temporal data. In the cross-modal attention fusion layer, the model additionally learns the deep semantic associations between salinity data and crop image stress features and soil moisture data.

[0116] This embodiment provides a detailed description of the expansion of the meteorological-water demand coupling analysis module:

[0117] During training, the deep spatiotemporal graph neural network of this module has two learning objectives:

[0118] Main output: Net irrigation requirement (Inet) of crops, which is the amount of water required to meet the evapotranspiration and growth of crops.

[0119] New auxiliary output: Based on the principle of salt balance, the leaching water requirement suggestion (Lrsuggest) is inferred by the network. This suggested amount is derived by the network based on the current root zone salt concentration, the salt tolerance threshold of the crop at the current growth stage, the mineralization of irrigation water, and the expected precipitation. It aims to keep the salt concentration in the root zone below a safe level.

[0120] Therefore, each irrigation unit in this irrigation district exist Total irrigation demand for the period Updated to:

[0121] ;

[0122] Inet represents the physiological water requirement. The definition of water requirement for leaching recommended by the network unifies the water requirements for crop physiological functions with the water requirements for soil environmental regulation.

[0123] This embodiment describes the extension of the channel system response modeling module with embedded physical constraints:

[0124] When abstracting the canal system into a directed graph, not only the water flow path is considered, but also the water quality (mainly mineralization) attribute is assigned to the water conveyance path from different water sources (such as reservoir water with low mineralization and well water with high mineralization) to different irrigation units.

[0125] In the multi-task loss function for model training, besides the original water balance physical residual term ( A new physical residual term for salt balance has been added. This term, based on a simplified convection-diffusion equation, calculates the difference between the network's predicted salt transport results and the physical laws, and is embedded into the training process as a soft constraint. This enables the trained model not only to predict water flow, but also to predict the changes in salinity (mineralization) as water is transported to each irrigation unit under different water distribution schemes.

[0126] Furthermore, this embodiment describes the expansion of the multi-objective collaborative optimization decision-making module:

[0127] Based on the three objectives of Example 1 (water demand satisfaction, smooth action, and safety constraints), a fourth objective is added: salinity control. A new reward function is also introduced. Designed as follows:

[0128] ;

[0129] This refers to the multi-objective reward in Example 1.

[0130] Fourth item (Salt coercion penalty item): It is the root zone soil salinity concentration of irrigation unit j in the next time period, predicted by the extended physical constraint embedding module after the current water distribution action is performed. It is the salt tolerance threshold of the crop currently planted in this unit at the current growth stage; the ReLU function ensures that a penalty is only incurred when the predicted salt content exceeds the tolerance threshold; It is its weighting coefficient.

[0131] During training, the strategy network must simultaneously learn how to dynamically coordinate four often conflicting objectives: water supply security, energy efficiency, system safety, and soil salinity control. For example, during periods of irrigation water scarcity, the strategy network may need to make difficult trade-offs: whether to prioritize meeting the physiological water needs of a cotton field or to prioritize leaching irrigation of a tomato field where salinity is nearing a critical threshold to prevent yield reduction due to salt damage. A well-trained strategy can generate intelligent water and salt joint regulation schemes.

[0132] Furthermore, this embodiment describes the expansion of the instruction execution and adaptive feedback correction module:

[0133] Drift detection conditions expanded: In addition to monitoring deviations in meteorological and hydraulic responses, the drift detector now uses an unexpected rate of increase in measured soil salinity as a trigger condition. When the rate of salinity accumulation is too rapid, exceeding the model's original understanding, it is determined that the environment has changed.

[0134] In the online fine-tuning process, in addition to using the latest water level, flow rate, and meteorological data, the latest soil salinity monitoring data will also be used to make targeted fine-tuning of relevant network parameters involved in salinity prediction, so that the system can adapt to the dynamic changes in soil salinity more quickly.

[0135] Furthermore, this embodiment describes the expansion of the credible decision interpretation and visualization module:

[0136] This module specifically enhances the analytical layers and interpretations related to salinity, for example:

[0137] An animation that dynamically displays the spatiotemporal evolution of soil salinity.

[0138] Predict and compare the overall salinity distribution of the irrigation area under different water source allocation schemes.

[0139] A trade-off analysis diagram is generated for the two major objectives of water conservation and salt control in the decision-making scheme, which intuitively shows the trade-offs between the two in this decision.

[0140] Through the implementation of this embodiment, the system of the present invention has successfully achieved intelligent and precise water-salt joint regulation in irrigation areas with extreme water scarcity and severe soil salinization. Based on real-time monitoring and multi-step prediction, the system can dynamically formulate optimized irrigation plans that meet the physiological water needs of crops while proactively preventing soil salinization. It can also timely and precisely schedule leaching irrigation before the crop's salt tolerance threshold. This achieves optimal spatial and temporal allocation of limited water resources (especially low-mineralized water resources), simultaneously achieving multiple benefits including water conservation, increased yield, salt control, and ecological environmental protection. It provides an innovative solution for the efficient, sustainable, and intelligent utilization of agricultural water resources in arid and semi-arid regions.

[0141] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0142] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An intelligent water distribution system for agricultural irrigation canals based on multimodal deep learning, characterized in that, include: The multimodal data fusion and representation module is used to acquire and fuse multi-source spatiotemporal sequence data from weather forecasts, soil moisture, crop growth remote sensing, and canal network operation status. It extracts unified feature representations with cross-modal semantic consistency through a pre-trained multimodal large model. The meteorological-water demand coupling analysis module is coupled to the multimodal data fusion and representation module. It receives the unified feature representations and models the nonlinear coupling relationship between meteorological factors, soil moisture dynamics, and crop evapotranspiration water consumption through a deep spatiotemporal graph neural network. It outputs refined water demand predictions for each irrigation zone in the future for multiple time periods. The canal system response modeling module with embedded physical constraints is used to construct a deep learning surrogate model with embedded physical priors based on the canal system topology, hydraulic parameters, and the law of conservation of mass and momentum. This model simulates the water level and flow propagation process within the canal system under different gate control commands. A multi-objective collaborative optimization decision-making module, connected to both the water demand prediction module and the canal system dynamic simulation module, is used to construct a sequential decision problem with water supply matching degree, operational energy efficiency, and safety boundary compliance as comprehensive optimization objectives. This problem is solved online using an agent or model predictive controller based on deep reinforcement learning to generate a forward-looking optimal water allocation command sequence. A command execution and adaptive feedback correction module, connected to the optimization decision-making module, is used to distribute the optimal water allocation command sequence to the canal system actuators and collect system status feedback and environmental observation data in real time. Through an online learning algorithm, the module dynamically updates the model parameters of the water demand prediction module and the canal system dynamic simulation module using the feedback and observation data.

2. The system according to claim 1, characterized in that, The multimodal large model adopts an attention mechanism-based architecture, which includes a visual encoder for processing image data, a temporal encoder for processing temporal data, and an encoder for processing structured metadata. The features output by each encoder interact and align through a cross-modal attention mechanism to achieve deep fusion of multi-source information.

3. The system according to claim 1, characterized in that, The graph structure of the deep spatiotemporal graph neural network model uses irrigation units as nodes, and the connection relationship between nodes is based on at least one of spatial proximity, soil property similarity and hydrological connectivity. The model captures spatial correlation through graph convolution operations and dynamically learns the influence weights of different input features on water demand prediction through an attention mechanism.

4. The system according to claim 1, characterized in that, The physical constraint-embedded canal system response modeling module abstracts the canal system into a directed graph and constructs a physical information neural network. During the forward propagation process, the network forces the nodes to meet the water balance through the embedded physical conservation layer and simulates the water flow resistance effect through a learnable momentum approximation layer, ensuring that the model output conforms to the basic laws of hydraulics.

5. The system according to claim 1, characterized in that, The multi-objective collaborative optimization decision module models the water allocation process as a constrained Markov decision process; its reward function is a multi-objective weighted sum, which includes at least: a weighted water demand deviation penalty term related to the priority of irrigation units, an action smoothing penalty term to suppress frequent fluctuations in control commands, and a physical constraint violation penalty term for water level or flow exceeding the safety limit.

6. The system according to claim 1, characterized in that, The instruction execution and adaptive feedback correction module includes a drift detection unit, which is used to continuously compare the difference between the predicted data and the actual monitoring data. When the difference is detected to continuously exceed the set threshold, an online learning process is triggered to fine-tune and update some parameters in the water demand prediction module and the canal system dynamic simulation module using recent data.

7. The system according to claim 1, characterized in that, It also includes a credible decision interpretation and visualization module, which is used to analyze and visualize the feature attention weights in the multimodal data fusion and representation module, the feature contribution in the water demand prediction module, and the action value assessment in the optimization decision module, and generate an explanatory report describing the decision basis and the multi-objective trade-off relationship.

8. A smart water distribution method for agricultural canal systems based on multimodal deep learning, characterized in that, include: By integrating weather forecasts, soil moisture data, crop remote sensing images, and irrigation system monitoring data, cross-modal features are extracted through a multimodal large model to form a unified feature representation; Meteorological-driven water demand prediction steps: Based on the unified feature representation, a deep spatiotemporal neural network model is used to analyze the coupling relationship between meteorology, soil, and crop systems, and predict the net irrigation water demand of each irrigation unit in the future for multiple time periods; Physically constrained canal system simulation steps: Based on the physical structure of the canal system and the laws of hydraulics, a deep learning proxy model that integrates physical information is constructed to predict the canal system state changes caused by water allocation commands; Multi-objective collaborative optimization decision-making steps: With the goal of minimizing water demand deviation, water conveyance energy consumption, and violation of physical constraints, under the constraints of the proxy model, the optimal water allocation command sequence is solved through deep reinforcement learning or model predictive control methods; Closed-loop execution and adaptive learning steps: The optimal water allocation command is executed, and real-time feedback data is collected. The parameters of the water demand prediction model and the canal system simulation model are dynamically updated through an online learning algorithm.

9. The method according to claim 8, characterized in that, In the multi-objective collaborative optimization decision-making step, the merits of the decision are evaluated by constructing a multi-objective reward function; the reward function includes at least a deviation penalty term reflecting the degree of water demand satisfaction, a smoothing penalty term reflecting control stability, and a constraint violation penalty term reflecting the safe operation of the system.

10. The method according to claim 8 or 9, characterized in that, In the closed-loop execution and adaptive learning steps, changes in the system or environment are identified by monitoring the continuous deviation between the predicted and actual values; when significant changes are identified, the water demand prediction model and the canal system simulation model are subjected to online incremental learning and parameter adjustment using the latest collected data sequence.

Images