Large-scale irrigation control device and method based on meteorological data
By using a large-scale irrigation control device based on meteorological data, and employing ESP32 and LoRa technologies for data acquisition and transmission, combined with a lightweight large-scale model and dynamic sparsity mechanism for water demand prediction, the data dependency and deployment cost issues of existing smart irrigation systems are resolved, achieving efficient and precise irrigation control.
Patent Information
- Application Number
- CN202611138334.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-11-10
AI Technical Summary
Existing smart irrigation systems have shortcomings in terms of data dependence, model deployment cost, short-term prediction accuracy, and control logic simplification. They are unable to meet the low-cost deployment needs of small and medium-sized planting areas and cannot achieve precise matching between irrigation volume and real-time crop water requirements.
A large-scale irrigation control device based on meteorological data is adopted, including a sensing layer, a transmission layer, a decision layer, and an execution layer. It uses ESP32 and LoRa technologies for data acquisition and transmission, combines a lightweight large-scale model and a dynamic sparsity mechanism for water demand prediction, eliminates feedback status data, and achieves precision irrigation through adaptive pump stations and irrigation valves.
It reduces data acquisition and processing costs, supports offline operation, improves prediction accuracy and system operating efficiency, simplifies control logic, reduces deployment and operation costs, and improves water resource utilization.
Smart Images

Figure CN123058239A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of irrigation technology, and more specifically, to a large-scale irrigation control device and method based on meteorological data. Background Technology
[0002] Agricultural irrigation is a core element in ensuring stable crop yields and increasing income, as well as improving water resource utilization efficiency. With the continuous advancement of high-standard farmland construction, traditional extensive irrigation methods are no longer adequate for the refined and water-saving development needs of modern agriculture. Problems such as untimely or excessive irrigation not only affect normal crop growth but also significantly increase production costs and cause serious water waste. In recent years, the deep penetration of artificial intelligence and Internet of Things technologies into agriculture has prompted intelligent irrigation systems to gradually replace traditional irrigation methods, becoming an important path to solving the aforementioned dilemmas.
[0003] Current smart irrigation technologies still face several pressing technical challenges. First, they rely on cumbersome and costly data. Most existing solutions depend on long-term historical meteorological data, remote sensing data, and crop growth cycle data for water demand forecasting and irrigation decisions. Acquiring, integrating, and storing this multi-dimensional, long-term data involves complex processes and high hardware and software costs, making it difficult to meet the low-cost deployment needs of small- to medium-sized planting areas. Furthermore, data update delays can lead to a disconnect between predicted water demand and actual crop water requirements, directly impacting irrigation accuracy. Second, large-scale model deployment has poor applicability and high costs. Existing solutions often use general-purpose large-scale model architectures, lacking lightweight adaptation and optimization for the specific scenario of short-term agricultural water demand forecasting. Deployment requires high-performance cloud computing power, significantly increasing system operating costs and posing security risks such as inability to function properly during network outages and leakage of agricultural data privacy during data transmission. Third, short-term water demand forecasting accuracy is insufficient. Most existing solutions rely solely on simple calculations based on meteorological data at a single moment, failing to deeply explore the intrinsic correlation between the temporal changes in meteorology over the past week (such as temperature trends, precipitation distribution, and accumulated sunshine) and the short-term water requirements of crops. This results in irrigation plans with low targeting and accuracy, making it difficult to accurately match irrigation volume with real-time crop water needs and fully realize the technological benefits of water-saving irrigation. Fourth, the system control logic is cumbersome and deployment costs are high. Some intelligent irrigation systems overly rely on execution layer feedback status data for parameter adjustment and closed-loop control, leading to complex system architecture, low response efficiency, and increased field equipment deployment and subsequent maintenance costs.
[0004] Therefore, existing intelligent irrigation systems have significant shortcomings in terms of data dependence, model deployment cost, short-term prediction accuracy, and simplified control logic. There is an urgent need for a new irrigation control scheme that can balance low-cost deployment, high-precision prediction, and simplified control logic to meet the actual application needs of intelligent and precise irrigation in planting areas of different sizes. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to provide a large-scale irrigation control device and method based on meteorological data.
[0006] Firstly, a large-scale irrigation control device based on meteorological data is provided, comprising a sensing layer, a transmission layer, a decision layer, and an execution layer that are sequentially connected in communication.
[0007] The perception layer includes a wireless meteorological interface module for acquiring and caching meteorological data;
[0008] The transmission layer includes edge computing nodes and a communication gateway. The edge computing nodes are used to preprocess the meteorological data to generate a standardized meteorological time-series dataset, and the communication gateway is used to realize data uplink transmission and command downlink transmission.
[0009] The decision-making layer includes a short-term meteorological large model module, a water demand prediction module, and an irrigation scheme optimization module, which are used to predict crop water demand based on the meteorological time series dataset, generate irrigation schemes, and output control commands.
[0010] The execution layer includes an irrigation execution mechanism for receiving and executing the control commands to complete the irrigation operation;
[0011] The device employs an open-loop control architecture, where the execution layer does not provide operational status data to the decision-making layer or the perception layer after performing irrigation operations.
[0012] Preferably, the wireless weather interface module includes an ESP32 control unit and is extended with a storage module via an SPI interface for caching and updating complete time-series weather data for the past week in real time. The wireless weather interface module is also configured with an interface failure backup mechanism, which enables local temporary caching in case of failure and automatically synchronizes and completes the data after the interface is restored.
[0013] Preferably, the edge computing node is an ESP32 module or an ARM processor. The edge computing node uses a filtering algorithm to remove outliers in the meteorological data and uses interpolation to complete the missing meteorological time series data.
[0014] Preferably, the communication gateway is a gateway with an integrated LoRa module, used to build a LoRa star network and support offline caching.
[0015] Preferably, the short-term meteorological large model module is also equipped with a dynamic sparsity mechanism and a role-specific adapter. The dynamic sparsity mechanism is used to dynamically activate core parameters and shield redundant parameters during model inference, and the role-specific adapter is used to output adapted time-series features for different crop types.
[0016] Preferably, the sparsity threshold of the dynamic sparsity mechanism adopts a two-level adjustment mechanism: the basic sparsity threshold is determined by using the comprehensive fluctuation amplitude of meteorological time series data as the main adjustment parameter, and the basic sparsity threshold is fine-tuned by using the hardware resource occupancy status of edge devices as an auxiliary constraint parameter.
[0017] Preferably, the irrigation scheme optimization module uses a reinforcement learning algorithm to simulate the soil moisture transport process under different irrigation parameters, optimizes the optimal irrigation scheme, and converts it into standardized control commands.
[0018] Preferably, the execution layer includes an adaptive pump station controller and a watering valve. The adaptive pump station controller is used to adjust the output power of the pump station, and the watering valve is used to control the valve's on / off state, on / off duration, and opening degree.
[0019] Secondly, a large-scale model irrigation control method based on meteorological data is provided, executed by any of the devices described in the first aspect, including:
[0020] S1. Collect meteorological data and preprocess it to generate a standardized meteorological time-series dataset;
[0021] S2. Load the lightweight large model and complete the initialization configuration;
[0022] S3. Convert the meteorological time-series dataset into a time-series feature vector, and extract meteorological time-series change features through the lightweight large model;
[0023] S4. Based on the meteorological time series change characteristics, call the crop short-term water response function library to predict crop water demand and key water demand time nodes;
[0024] S5. Based on the water demand prediction results, generate the optimal irrigation plan through an optimization algorithm;
[0025] S6. Convert the optimal irrigation scheme into control commands and send them to the execution layer to perform irrigation operations.
[0026] Preferably, in step S3, the conversion method of the meteorological time series data into time series feature vector is a learnable linear mapping combined with position encoding; in step S4, after generating the prediction result, the water demand prediction module calls the verification dataset to verify the prediction result. If the error exceeds the preset threshold, the weights are adjusted in reverse and recalculated.
[0027] The beneficial effects of this invention are:
[0028] 1. This invention adopts a lightweight mode of "last week's meteorological time series data", which eliminates the need for soil data, feedback data and complex multi-dimensional data, greatly reducing the cost of data acquisition and processing; it adopts a wireless meteorological interface, eliminating the need for wiring, further reducing the difficulty and cost of field deployment, while ensuring the integrity of short-term meteorological data through the ESP32 storage module.
[0029] 2. The dynamic sparsity mechanism and role-specific adapter of this invention work together to achieve lightweight edge deployment of large models, balancing computing power saving and prediction accuracy, supporting offline operation, ensuring data privacy and real-time response, and eliminating the need to rely on feedback data to optimize model parameters.
[0030] 3. The ESP32 of this invention is fully adaptable to small-scale scenarios, allows for flexible selection of edge computing nodes, and LoRa low-power communication is adapted to field scenarios. It works in conjunction with the wireless meteorological interface of the perception layer to reduce deployment and operation costs. The wireless meteorological interface provides stable transmission and improves the reliability of data reception. The execution layer eliminates the gate group controller and achieves irrigation regulation through precise pressure control of pump stations combined with precise flow control of irrigation valves. At the same time, the status feedback function is eliminated, simplifying the architecture and reducing deployment costs and communication complexity.
[0031] 4. This invention simplifies the “collection-decision-execution” link, eliminates the need for feedback status data, reduces data transmission pressure and equipment response burden, improves system operating efficiency, and enhances water resource utilization through adaptive pump station and watering valve coordinated control. Attached Figure Description
[0032] Figure 1 The structural block diagram of the large-scale irrigation control device based on meteorological data provided by the present invention;
[0033] Figure 2 The flowchart of the large-scale irrigation control method based on meteorological data provided by the present invention. Detailed Implementation
[0034] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
[0035] Example 1:
[0036] To address the problems of the prior art, Embodiment 1 of this application provides a large-scale irrigation control device based on meteorological data, comprising: a sensing layer, a transmission layer, a decision layer, and an execution layer that are sequentially connected in communication.
[0037] The perception layer includes a wireless meteorological interface module, which is used to acquire meteorological data and cache it.
[0038] The wireless weather interface module includes an ESP32 control unit and a storage module extended via an SPI interface for caching and updating complete time-series weather data for the past week. The wireless weather interface module is also configured with an interface failure backup mechanism, which activates the local temporary cache in case of failure and automatically synchronizes and completes the data after the interface recovers.
[0039] The transmission layer includes edge computing nodes and a communication gateway. The edge computing nodes are used to preprocess the meteorological data to generate a standardized meteorological time-series dataset, and the communication gateway is used to realize data uplink transmission and command downlink transmission.
[0040] The edge computing node is an ESP32 module or an ARM processor. The edge computing node uses a filtering algorithm to remove outliers in the meteorological data and uses interpolation to complete the missing meteorological time series data.
[0041] The communication gateway is a gateway with an integrated LoRa module, used to build a LoRa star network and supports offline caching.
[0042] The decision-making layer includes a short-term meteorological model module, a water demand prediction module, and an irrigation scheme optimization module, which are used to predict crop water demand based on the meteorological time series dataset, generate irrigation schemes, and output control commands.
[0043] The short-term meteorological large model module is also equipped with a dynamic sparsity mechanism and a role-specific adapter. The dynamic sparsity mechanism is used to dynamically activate core parameters and shield redundant parameters during model inference. The role-specific adapter is used to output adapted time-series features for different crop types.
[0044] The sparsity threshold of the dynamic sparsity mechanism adopts a two-level adjustment mechanism: the comprehensive fluctuation amplitude of meteorological time series data is used as the main adjustment parameter to determine the basic sparsity threshold, and the hardware resource occupancy status of edge devices is used as an auxiliary constraint parameter to fine-tune the basic sparsity threshold.
[0045] The irrigation scheme optimization module uses a reinforcement learning algorithm to simulate the soil moisture transport process under different irrigation parameters, optimizes the optimal irrigation scheme, and converts it into standardized control commands.
[0046] The execution layer includes an irrigation execution mechanism for receiving and executing the control commands to complete the irrigation operation.
[0047] The execution layer includes an adaptive pump station controller and a watering valve. The adaptive pump station controller is used to adjust the output power of the pump station, and the watering valve is used to control the valve's on / off state, on / off duration, and opening degree.
[0048] The device employs an open-loop control architecture, where the execution layer does not provide operational status data to the decision-making layer or the perception layer after performing irrigation operations.
[0049] Example 2:
[0050] Building upon Example 1, Example 2 of this application provides a more specific large-scale irrigation control device based on meteorological data. Driven by "deep mining of short-term meteorological data," supported by Alibaba's Qwen-1.8B open-source platform, and guaranteed by "low-cost hierarchical architecture and simplified control logic," it eliminates the need for long-term historical meteorological data, remote sensing data, crop growth cycle data, or other complex data. It does not rely on feedback status data for adjustment; instead, it utilizes only the meteorological time-series data from the past week, combined with the time-series feature extraction and water demand prediction capabilities of the Qwen-1.8B open-source lightweight large-scale model, along with LoRa low-power communication and edge computing technologies, to achieve precise generation and efficient execution of irrigation plans, adapting to the needs of planting areas of different sizes.
[0051] Specifically, such as Figure 1 As shown, the device includes a perception layer, a transmission layer, a decision layer, and an execution layer that are connected in sequence, realizing end-to-end coordination of data acquisition, transmission, decision-making, and execution, simplifying control logic, and eliminating the need for feedback status data.
[0052] The perception layer is the meteorological data acquisition terminal, used to provide raw meteorological data support. Its core functions are: acquiring hourly meteorological data for the past week and current meteorological data through a wireless meteorological interface; performing preliminary data format conversion and caching; and providing high-quality raw meteorological data for subsequent preprocessing. It does not need to collect soil-related data; it is only responsible for data acquisition and uploading and does not participate in status feedback. The hardware selection and functions of the perception layer are as follows:
[0053] Core module: Meteorological interface module
[0054] Hardware configuration: The core control unit is ESP32, which integrates a wireless weather interface and is deployed in an unobstructed location at the edge of the planting area to ensure stable reception of weather data;
[0055] Auxiliary functions: Real-time access to meteorological data (temperature, rainfall, light intensity, air humidity) via wireless meteorological interface, initial data format conversion, and storage module expansion via SPI interface (such as 16GB TF card) to cache complete time-series meteorological data for the past week and update it in real time to ensure data continuity; Supports interface failure backup mechanism, which enables local temporary caching in case of failure, and automatically synchronizes and completes data after the interface is restored.
[0056] The transport layer is the data and command transmission hub, used for efficient preprocessing and transmission. Its core functions are: receiving raw meteorological data from the sensing layer, performing preprocessing such as data cleaning, outlier removal, and time series completion, and generating a standardized meteorological time series dataset for the past week; simultaneously enabling data uplink transmission from the sensing layer to the decision layer and command downlink transmission from the decision layer to the execution layer, adapting to the low-power, long-distance communication requirements in the field, and only responsible for data and command transmission, without transmitting feedback data. The hardware selection and functions of the transport layer are as follows:
[0057] Core Module 1: Edge Computing Node
[0058] Hardware selection: Flexible adaptation according to planting scale - small-scale planting areas (within 5 acres) use ESP32-WROOM-32E modules, while large-scale planting areas use ARM Cortex-A72 processors;
[0059] Preprocessing function: Running the FreeRTOS operating system, the Kalman filter algorithm is used to remove outliers in meteorological data, and the linear interpolation method is used to complete the missing meteorological time series data. The preprocessing time for a single set of data does not exceed 800ms (of which the Kalman filter time is ≤300ms), and a standardized meteorological time series dataset of the past week is generated.
[0060] Advantages: The ESP32 version consumes only 1 / 3 the power of traditional industrial-grade chips, making it suitable for solar power supply scenarios in the field and enabling efficient data transmission in collaboration with the wireless meteorological interface of the sensing layer.
[0061] Core Module 2: Communication Gateway
[0062] Hardware selection: ESP32 gateway with integrated LoRa module (small-scale scenario) or industrial-grade LoRa gateway (large-scale scenario).
[0063] Communication Functions: Construct a LoRa star network to achieve short-range, low-speed communication with the wireless meteorological interface of the sensing layer and edge computing nodes, complete the uplink transmission of meteorological data and the downlink transmission of commands, support network outage caching, and update data every 30 minutes.
[0064] The decision-making layer is the core of the intelligent system, used to achieve precise decision-making and scheme optimization. Its core functions are: relying on a large-scale agricultural model, it extracts meteorological time-series features over the past week, predicts crop water requirements for the next 48 hours, generates and optimizes irrigation schemes based on irrigation district data, and outputs standardized control commands. It is the "decision brain" of the entire system, without relying on soil data or status data from the execution layer for prediction and decision-making. The hardware selection and functions of the decision-making layer are as follows:
[0065] Core Module 1: Short-Term Weather Model Module
[0066] Model selection: Qwen-1.8B, an open-source, lightweight large model, deployed on a local edge server (not in the cloud).
[0067] Hardware support: The edge server is equipped with an Intel Xeon Gold processor and 64GB of memory, and is paired with the "Shennong-Tuojiang" domestic computing power base, which supports offline operation.
[0068] Core Technology: Employing a dynamic sparsity mechanism (dynamically adjusting the sparsity threshold from 30% to 60%) and role-specific adapter technology, and based on an optimized Transformer time-series modeling architecture, this system dynamically activates core parameters and masks redundant parameters, reducing computational power consumption by 50%. The sparsity threshold uses a two-tiered adjustment mechanism: "business parameters as the primary factor, hardware parameters as the secondary factor." The primary adjustment parameter is the comprehensive fluctuation range of multi-dimensional meteorological time-series data (temperature, humidity, light intensity, rainfall, and wind speed) over the past 24 hours. A weighted comprehensive fluctuation score is obtained by calculating the coefficient of variation of each meteorological indicator, and a basic sparsity threshold within the 30%–60% range is determined based on this score. The greater the meteorological time-series fluctuation, the lower the set sparsity threshold, thus ensuring the accuracy of crop water requirement prediction under extreme weather conditions. Simultaneously, the real-time CPU idle rate and available memory ratio of edge devices are used as auxiliary constraint parameters to fine-tune the basic sparsity threshold upwards. When device hardware resources are strained, the sparsity is appropriately increased to reduce model computational overhead, and the final sparsity is strictly limited to 30%–60%. Within the specified range, the threshold adjustment range between two consecutive adjustments does not exceed ±10%, ensuring stable operation of edge devices while maintaining prediction accuracy.
[0069] The role-specific adapter mounted in this application is the LoRA (Low-Rank Adapter) network structure. All matrix parameters of the LoRA adapter remain densely activated throughout the process and do not participate in the dynamic sparse masking in the 30%~60% interval. Independent A and B weight parameter sets are trained for different crops such as corn, wheat, fruits and vegetables. By simply switching the adapter parameters, the output can be customized with the specific time-series features of the corresponding crop growth period under the same meteorological time-series input, thus achieving crop role-specific adaptation.
[0070] Model Functionality: Extracts meteorological time-series features (temperature rise and fall trends, precipitation distribution, cumulative sunshine, etc.) over the past week, providing core feature support for water demand forecasting, with inference latency controlled within 200ms and water demand forecasting error ≤ ±4%.
[0071] Training details: The training set is "weather time series data of the past week - measured data of crop water requirements". A sparse regularization term is introduced for optimization. The batch size is 32 and the initial learning rate is 0.001. The parameters are optimized through cross-validation.
[0072] Core Module 2: Water Demand Forecasting Module
[0073] The function of the water demand prediction module is to predict crop water demand and key water demand time nodes in the next 48 hours based on meteorological features extracted from a large model and combined with the built-in crop short-term water response function library (adapted to different crops), without the need for soil moisture data and execution layer feedback data.
[0074] Core Module 3: Irrigation Scheme Optimization Module
[0075] Function: Based on water demand prediction results, generate an initial irrigation plan (including irrigation time and irrigation amount); simulate the soil moisture transport process through reinforcement learning algorithm, optimize to obtain the optimal plan including irrigation control parameters, and convert it into standardized control instructions without referring to the execution layer feedback status data.
[0076] In the decision-making level, the complete operational mechanism of time-series feature extraction and water demand forecasting results is the core of achieving accurate decision-making. It consists of four key stages, which are explained in detail below:
[0077] 1. Phase One: Data Input and Model Initialization (Connecting to the Transport Layer)
[0078] For water demand forecasting to be accurate, data input and model initialization must be completed first to lay the foundation for time series feature extraction and prediction. This includes the following steps:
[0079] 1.1 Data Input: After preprocessing, the edge computing node of the transmission layer transmits the "standardized meteorological time series dataset of the past week" (including hourly temperature, precipitation, sunshine, wind speed and air humidity data) to the short-term meteorological large model module of the decision layer through the LoRa communication gateway. No soil-related data or any feedback data need to be input. This dataset comes from the raw meteorological data collected by the wireless meteorological interface of the perception layer. After preprocessing, the data integrity and standardization are ensured.
[0080] 1.2 Model Initialization: The decision layer loads the trained Qwen 1.8B open-source lightweight large model and configures the core parameters through low-code tools—setting the time window length to 7 days × 24 hours (covering hourly data for nearly a week), strengthening the attention weights of temperature and precipitation features, and enabling the corn (or corresponding crop) dedicated adapter to ensure that the model is adapted to the current planting scenario.
[0081] It should be noted that the model has been optimized in advance using the training set of "weather time series data of the past week - measured crop water requirement data". The dynamic sparsity mechanism is enabled by default, and the initial sparsity threshold is set to 45% (only 55% of the core parameters are activated). This balances computational power consumption and prediction accuracy, and the inference latency is controlled within 200ms, meeting the requirements for real-time operation without relying on feedback data to optimize parameters.
[0082] 2. Phase Two: Temporal Feature Extraction (led by the short-term meteorological large model module)
[0083] Temporal feature extraction is a core prerequisite for water demand forecasting. It is independently completed by the short-term meteorological model module. Its core is to mine trend features that are strongly correlated with crop water demand from meteorological data of the past week, rather than simply extracting a single data point. It is completed solely based on the input meteorological data and does not receive any feedback data. The specific operation is as follows:
[0084] 2.1 Data Format Conversion: The large model first converts the standardized meteorological time-series dataset of the past week into a 128-dimensional time-series feature vector that it can recognize (as configured in Example 1) to ensure that the data matches the model input format. This dataset comes from hourly meteorological data of the past week collected by the wireless meteorological interface of the perception layer. In this scheme, the conversion of meteorological time-series data into 128-dimensional time-series feature vectors is achieved by learningable linear embedding fully connected mapping + sine and cosine position encoding.
[0085] 2.2 Feature Depth Mining: The model uses the Transformer temporal attention mechanism to analyze the 128-dimensional feature vector, focusing on extracting three core temporal features (directly related to crop water requirements):
[0086] 2.3 Temperature rise / fall trend characteristics: such as the rate of change of the daily average temperature from 18℃ to 28℃ in the past week, and the cumulative duration of high temperature period (≥30℃), reflecting the change in crop transpiration intensity;
[0087] 2.4 Sparse precipitation distribution characteristics: such as the distribution of precipitation periods in the past week, cumulative rainfall, and the duration of continuous precipitation, to determine the soil moisture replenishment;
[0088] 2.5. Characteristics of cumulative light: such as the average daily light intensity (3000-9000 lux) and total cumulative light over the past week, reflecting the water requirements of crop photosynthesis.
[0089] 2.6 Feature Selection and Optimization: The model utilizes a built-in role-specific adapter, a lightweight, plug-and-play module designed specifically for short-term agricultural water demand forecasting. Mounted on the Qwen1.8B large model's Transformer layer, it achieves low-cost adaptation from a general-purpose large model to an agricultural-specific model with minimal parameter increments, without modifying the base weights. This strengthens the correlation between meteorological time-series features and crop water demand, while shielding redundant features unrelated to agricultural water demand (such as extreme wind speed anomalies). Only the three core features mentioned above are retained. The dynamic sparsity mechanism ensures that the large model only activates core parameters related to "agricultural meteorological water demand forecasting" during inference, shielding useless parameters. This achieves halved computing power, doubled speed, and no decrease in accuracy. Real-time adjustment of parameter activation states ensures the accuracy and efficiency of feature extraction, ultimately outputting a "core time-series feature set."
[0090] For example, in a corn-growing area scenario, the wireless meteorological interface of the perception layer collects hourly meteorological data for the past week. After preprocessing by the transmission layer, the data is input into a large model. The model extracts the time-series characteristics of "continuous temperature rise in the past week, cumulative precipitation of 12 mm, 48 consecutive hours without precipitation, and sufficient cumulative sunshine," providing core basis for subsequent water demand forecasting. No feedback data was received throughout the process.
[0091] 3. Phase Three: Generation of Water Demand Forecast Results (led by the water demand forecast module)
[0092] The water demand forecasting module receives the "core time-series feature set" output by the short-term meteorological big data model module. Combined with built-in resources, it completes crop water demand forecasting. The core is to transform meteorological time-series features into quantifiable and applicable forecast results without the need for soil data or feedback data. The specific operation is as follows:
[0093] 3.1 Feature and Water Demand Correlation Matching: The water demand prediction module calls the built-in "Crop Short-Term Water Response Function Library". The crop short-term water response function is a mapping model for predicting crop water demand for the next 48 hours based on the meteorological time series of the past week for four meteorological factors: temperature, humidity, light intensity, and precipitation. Input: hourly meteorological time series of the past week: temperature T(t), relative humidity RH(t), light intensity I(t), precipitation R(t). Construct the baseline water demand model ET0(t), introduce the crop-growth period coefficient Kc(τ), and construct the short-term water demand integral function, the formula is:
[0094] W 48 (τ) = Kc(τ) × ∫0 48 ET0(t) · λ(t) dt
[0095] Among them, W 48Let Kc(τ) be the water requirement for the next 48 hours, Kc(τ) be the crop and growth period coefficient, ET0(t) be the baseline water requirement, and λ(t) be the weighted coefficient of the meteorological time series over the past week. The formula for calculating ET0(t) is:
[0096] ET0(t) = α·T(t) + β·(1-RH(t)) + γ·I(t) - ε·R(t)
[0097] Where T is temperature, 1-RH is humidity, I is light intensity, and R is precipitation (negative inhibition).
[0098] Furthermore, λ(t) and W 48 The formula for calculating _corn(τ) is:
[0099] λ(t) = 0.65·T_trend + 0.25·R_deficit + 0.10·I_sum
[0100] W 48 _corn(τ) = Kc_corn(τ) × ∫0 48 [αT + β(1-RH) + γI-εR] · λ(t) dt
[0101] It should be noted that the formula α, β, γ, and ε are empirical constants calibrated for different regions and climate scenarios, not dynamic weights directly output by the large model. The dynamic weights output by the large model are λ(t) (the time-series weighting coefficient of meteorological data for the past week) in the integral term. The two have different functions.
[0102] Specifically, α, β, γ, and ε characterize the inherent physical contribution of meteorological factors to baseline evapotranspiration at a single moment. They are obtained by linear regression fitting of multi-year field observation data in the region and are reused for a long time in the same climate zone and for the same type of crop. These are the time-varying dynamic weights output by the Transformer time-series model, depicting the lagging trend of weather fluctuations over the past 7 days and their impact on the hourly water demand in the future. Each prediction window changes dynamically with historical time series. Based on the current crop type (e.g., corn), the corresponding response function is matched (e.g., the correlation function between corn's short-term water demand and temperature, light, and precipitation), establishing a mapping relationship between "time series characteristics → water demand".
[0103] Taking wheat and corn fields as examples, the coefficients obtained through multi-year observation and regression calibration are shown in Table 1.
[0104] Table 1
[0105]
[0106] 3.2 Quantitative Prediction Calculation: Based on the mapping relationship, the crop water demand in the next 48 hours is quantitatively predicted through function calculation, and key water demand time nodes are identified. Key nodes are selected first from the "time periods when meteorological characteristics have the greatest impact on water demand" (such as the time periods with the highest temperature and the strongest sunlight).
[0107] 3.3 Validation and Correction of Prediction Results: The validation dataset retained during model training is used to validate the preliminary prediction results, ensuring that the prediction error is controlled within ±4%. If the error exceeds the threshold, the time-series feature weights are adjusted in reverse, and the calculation is repeated until the accuracy requirements are met. The validation process does not rely on any real-time feedback data. The validation dataset used in this scheme is a historical static field observation dataset, not real-time collected data. It does not directly compare the "future real water demand of this farmland" in real time, but uses a historical similar sample matching and validation mechanism with the same scene, time period, and climate characteristics to achieve accuracy validation. This is an offline pre-constraint validation, which does not rely on the actual future measured feedback of the current plot throughout the entire process.
[0108] 3.4 Output standardized prediction results: Two types of core outputs are generated in the end, both of which are the direct basis for subsequent irrigation scheme optimization.
[0109] 3.5 Quantify water requirements: such as "the water requirement for corn in the next 48 hours is 60 m³ / mu".
[0110] 3.6 Key water demand points: such as "10:00 AM the next day" in Example 1 (the time when the temperature is higher and the water demand is most urgent).
[0111] It should be noted that the prediction results of this application do not rely on long-term historical meteorological data, soil data, or any feedback status data. They are based solely on the time-series meteorological data collected over the past week from the wireless meteorological interface of the sensing layer. This fully aligns with the core design of this patent, which is "lightweight data-driven and simplified control logic." At the same time, it supports dynamically adjusting the response function parameters according to the crop growth stage to adapt to the differences in water requirements at different growth stages.
[0112] 4. Phase Four: Application of Prediction Results (Linked Execution Layer).
[0113] Water demand forecasts are not output in isolation, but directly drive irrigation scheme optimization and are then distributed to the execution layer for implementation, simplifying control logic and eliminating the need for dynamic adjustments based on feedback data. The specific operation is as follows:
[0114] 4.1 Directly applied to irrigation scheme optimization: The water demand prediction module transmits the "quantified total water demand + key water demand nodes" of the current single green space to the irrigation scheme optimization module as the only core input; the irrigation scheme optimization module generates an initial irrigation scheme for a single plot based on the crop type and soil parameters (including the optimal irrigation time period and the total irrigation volume per irrigation).
[0115] 4.2 Command Issuance and Execution: The decision-making level converts the optimized single-plot irrigation plan into standardized control commands, which are then sent to the field pump stations and solenoid valve actuators through the transmission layer to directly carry out irrigation operations on the current entire crop area. The system does not need to receive any feedback data such as the operating status of the execution layer equipment or soil moisture conditions. This open-loop unidirectional control simplifies the operation logic, reduces the cost of software and hardware deployment, and improves the stability and efficiency of the system operation.
[0116] The execution layer is the action output terminal, used to implement irrigation plans. Its core function is to receive control commands from the decision-making layer, execute specific irrigation operations, and strictly follow the commands to complete the irrigation actions. It does not need to provide any operational status data back to the decision-making or perception layers, simplifying equipment logic and communication costs. The hardware selection and functions of the execution layer are as follows:
[0117] Hardware configuration: Integrated pressure sensor;
[0118] Function: Dynamically adjusts the output power of the pump station to ensure stable irrigation network pressure, accurately matches irrigation demand, avoids uneven irrigation, replaces traditional gate group control to achieve irrigation volume and regional adaptation, and does not require feedback of pressure, operating status and other data after executing the command.
[0119] Core module: Watering valve
[0120] Hardware configuration: Integrated electromagnetic control unit and flow sensing component;
[0121] Function: Based on irrigation control instructions issued by the decision-making level, it can precisely control the valve opening and closing status, opening duration and opening degree to achieve precise regulation of irrigation water volume. It can also work with the adaptive pump station controller to complete efficient irrigation. After executing the instructions, there is no need to provide feedback on valve status, flow rate and other data.
[0122] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.
[0123] Example 3:
[0124] Based on Example 2, Example 3 of this application provides a large-scale model irrigation control method based on meteorological data, which realizes efficient control across the entire chain of "collection-preprocessing-analysis-prediction-optimization-execution" without the need to collect soil-related data or feedback status data. The specific steps are as follows:
[0125] S1. Collect meteorological data and preprocess it to generate a standardized meteorological time-series dataset.
[0126] Specifically, the wireless meteorological interface module of the perception layer acquires hourly meteorological data (temperature, rainfall, light intensity, and air humidity) for the past week and meteorological data at the current moment, updates the data sequence and caches it; transmits all raw meteorological data to the edge computing node of the transmission layer, removes outliers in the meteorological data using the Kalman filter algorithm, and completes the missing time-series meteorological data using the linear interpolation method to generate a standardized meteorological time-series dataset for the past week.
[0127] S2. Load the lightweight large model and complete the initialization configuration.
[0128] Specifically, the decision layer loads the trained Qwen-1.8B open-source lightweight large model, initializes the model parameters through low-code tools, focuses on configuring the time window length (7 days × 24 hours), attention mechanism weights (strengthening temperature and precipitation feature weights), and configures the corresponding crop-specific adapter to ensure that the model is adapted to the current planting scenario.
[0129] S3. Convert the meteorological time series dataset into a time series feature vector, and extract the meteorological time series change features through the lightweight large model.
[0130] Specifically, the standardized meteorological time-series dataset of the past week is converted into a time-series feature vector (e.g., 128-dimensional) that the model can recognize, and input into the short-term meteorological large model module; the model uses a time-series attention mechanism to deeply extract the meteorological change trend features (temperature rising trend, sparse precipitation distribution features, cumulative sunshine features, etc.) within the past week, and completes the in-depth mining of short-term meteorological data.
[0131] S4. Based on the meteorological time series change characteristics, call the crop short-term water response function library to predict crop water demand and key water demand time nodes.
[0132] Specifically, the large model combines extracted meteorological trend features with a built-in crop short-term water response function library to infer the current crop water demand status. It can accurately predict crop water demand and key water demand time points in the next 48 hours using only meteorological features, without the need for soil data.
[0133] S5. Based on the water demand prediction results, generate the optimal irrigation plan through an optimization algorithm.
[0134] Specifically, based on the water demand prediction results, an initial irrigation plan (including irrigation time and irrigation volume) is generated; through reinforcement learning algorithms, the soil moisture transport process and crop growth status under different irrigation parameters are simulated, and the optimal irrigation plan including irrigation time, irrigation volume and irrigation control parameters is obtained.
[0135] S6. Convert the optimal irrigation scheme into control commands and send them to the execution layer to perform irrigation operations.
[0136] Specifically, the decision-making layer converts the optimal irrigation plan into standardized control commands, which are then sent to the execution layer through the communication gateway of the transmission layer. This controls the adaptive pumping station and watering valves to coordinate the irrigation operation. The execution layer strictly follows the commands to complete the irrigation action without needing to provide any operational status data back to the decision-making layer.
[0137] It should be noted that the method provided in this embodiment is the method corresponding to the device provided in embodiment 2. Therefore, the parts that are the same as or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be repeated in this application.
Claims
1. A large-scale irrigation control device based on meteorological data, characterized in that, It includes the perception layer, transmission layer, decision layer, and execution layer, which are connected in sequence via communication. The perception layer includes a wireless meteorological interface module for acquiring and caching meteorological data; The transmission layer includes edge computing nodes and a communication gateway. The edge computing nodes are used to preprocess the meteorological data to generate a standardized meteorological time-series dataset, and the communication gateway is used to realize data uplink transmission and command downlink transmission. The decision-making layer includes a short-term meteorological large model module, a water demand prediction module, and an irrigation scheme optimization module, which are used to predict crop water demand based on the meteorological time series dataset, generate irrigation schemes, and output control commands. The execution layer includes an irrigation execution mechanism for receiving and executing the control commands to complete the irrigation operation; The device employs an open-loop control architecture, where the execution layer does not provide operational status data to the decision-making layer or the perception layer after performing irrigation operations.
2. The large-scale irrigation control device based on meteorological data according to claim 1, characterized in that, The wireless weather interface module includes an ESP32 control unit and a storage module extended via an SPI interface for caching and updating complete time-series weather data for the past week. The wireless weather interface module is also configured with an interface failure backup mechanism, which activates the local temporary cache in case of failure and automatically synchronizes and completes the data after the interface recovers.
3. The large-scale irrigation control device based on meteorological data according to claim 2, characterized in that, The edge computing node is an ESP32 module or an ARM processor. The edge computing node uses a filtering algorithm to remove outliers in the meteorological data and uses interpolation to complete the missing meteorological time series data.
4. The large-scale irrigation control device based on meteorological data according to claim 3, characterized in that, The communication gateway is a gateway with an integrated LoRa module, used to build a LoRa star network and supports offline caching.
5. The large-scale irrigation control device based on meteorological data according to claim 4, characterized in that, The short-term meteorological large model module is also equipped with a dynamic sparsity mechanism and a role-specific adapter. The dynamic sparsity mechanism is used to dynamically activate core parameters and shield redundant parameters during model inference. The role-specific adapter is used to output adapted time-series features for different crop types.
6. The large-scale irrigation control device based on meteorological data according to claim 5, characterized in that, The sparsity threshold of the dynamic sparsity mechanism adopts a two-level adjustment mechanism: the comprehensive fluctuation amplitude of meteorological time series data is used as the main adjustment parameter to determine the basic sparsity threshold, and the hardware resource occupancy status of edge devices is used as an auxiliary constraint parameter to fine-tune the basic sparsity threshold.
7. The large-scale irrigation control device based on meteorological data according to claim 6, characterized in that, The irrigation scheme optimization module uses a reinforcement learning algorithm to simulate the soil moisture transport process under different irrigation parameters, optimizes the optimal irrigation scheme, and converts it into standardized control commands.
8. The large-scale irrigation control device based on meteorological data according to claim 7, characterized in that, The execution layer includes an adaptive pump station controller and a watering valve. The adaptive pump station controller is used to adjust the output power of the pump station, and the watering valve is used to control the valve's on / off state, on / off duration, and opening degree.
9. A large-scale model-based irrigation control method based on meteorological data, characterized in that, Performed by the apparatus according to any one of claims 1 to 8, comprising: S1. Collect meteorological data and preprocess it to generate a standardized meteorological time-series dataset; S2. Load the lightweight large model and complete the initialization configuration; S3. Convert the meteorological time-series dataset into a time-series feature vector, and extract meteorological time-series change features through the lightweight large model; S4. Based on the meteorological time series change characteristics, call the crop short-term water response function library to predict crop water demand and key water demand time nodes; S5. Based on the water demand prediction results, generate the optimal irrigation plan through an optimization algorithm; S6. Convert the optimal irrigation scheme into control commands and send them to the execution layer to perform irrigation operations.
10. The large-scale irrigation control method based on meteorological data according to claim 9, characterized in that, In step S3, the conversion method of the meteorological time series data into time series feature vector is a learnable linear mapping combined with position encoding; in step S4, after generating the prediction result, the water demand prediction module calls the verification dataset to verify the prediction result. If the error exceeds the preset threshold, the weights are adjusted in reverse and recalculated.
