A greening slope intelligent water-saving irrigation method and system based on multivariable decoupling and predictive control

Abstract

The application discloses a greening slope intelligent water-saving irrigation method and system based on multivariate decoupling and predictive control. The application divides the slope into N units, collects multi-source sensing data and pre-processes; constructs soil permeation factor Kr and hydraulic gradient I, decouples the water balance equation of multivariate coupling into one-dimensional equations of each unit; based on the decoupling model, a state space equation is established, a constrained objective function is optimized by MPC, and the current optimal valve opening, irrigation time and other control quantities are output; after execution, the prediction model is corrected online with actual feedback error to form a closed loop; the edge computing platform runs the above steps in real time, and the cloud training data drives the model to assist in adjusting the weight parameters. The application greatly reduces the complexity of multivariate coupling, compensates for the large hysteresis and runoff disturbance of the slope surface, realizes real-time water-saving irrigation on low-cost edge devices, and is suitable for unmanned management and protection of greening slopes of highways, mines, rivers and the like.

Description

Technical Field

[0001] This invention relates to the fields of slope ecological restoration, intelligent agricultural irrigation and automatic control technology, specifically a method and system for intelligent water-saving irrigation of green slopes based on multivariable decoupling and predictive control. Background Technology

[0002] Slope greening projects are typically located in remote mountainous areas, along roadsides, or abandoned mining sites, where power supply and water access are difficult, and often require unattended operation. Traditional irrigation control methods are mostly timed irrigation or on / off control based on single-point soil moisture thresholds, which are generally applicable on flat ground, but reveal serious shortcomings under slope conditions. First, slope irrigation exhibits significant gravity migration of water flow. Irrigation water from the upstream area flows downhill and diffuses to both sides, resulting in the upstream area being flooded while the downstream area remains undisturbed. Ineffective runoff carries away a large amount of valuable fertilizer, causing environmental pollution and economic losses. Second, slope irrigation systems exhibit typical characteristics of strong multivariate coupling, large lag, and nonlinearity. The soil infiltration rate in a single area is interrelated with more than 20 parameters, including soil texture, sensor burial depth, slope, water supply pressure, electric valve opening, and water supply from adjacent areas. Moreover, the time from irrigation start to sensor detection and then to moisture change often exceeds 30 minutes. This large lag makes conventional feedback control prone to overshoot and oscillation.

[0003] Currently, the commonly used proportional-integral-derivative (PID) controllers in engineering are simple in design, but are essentially single-input single-output (SISO) linear controllers, unable to handle multivariable coupling and large time delay problems. Some research has attempted to use fuzzy control or neural networks, but most remain at the simulation stage, and are computationally intensive and lack constraints on the specific physical mechanisms of slopes, such as failing to consider slope protection against soil erosion and fertility loss. Model predictive control (MPC) is theoretically suitable for multivariable constrained control, but classic MPC requires accurate mechanistic models and online solving of quadratic programming (QP) is computationally intensive, making it difficult to run in real time on low-cost edge controllers. Summary of the Invention

[0004] The purpose of this invention is to propose an intelligent water-saving irrigation method and system for green slopes based on multivariate decoupling and predictive control. This intelligent water-saving irrigation method can decouple multiple variables by utilizing the physical laws of slope irrigation and achieve rolling optimization through lightweight calculations, so as to meet the engineering needs of unattended maintenance, water and fertilizer conservation, and low-cost deployment of green slope plants in the later stage.

[0005] A smart water-saving irrigation method for green slopes based on multivariate decoupling and predictive control includes:

[0006] Data acquisition and preprocessing steps: Divide the slope into N units, collect multi-source sensor data from each unit and perform filtering and noise reduction, construct a sliding window dataset and unify the timestamps to obtain soil moisture after unifying the timestamps. Soil electrical conductivity Combined concentration of nitrogen, phosphorus and potassium Soil temperature Soil pH value Actual burial depth of the sensor Slope angle Opening degree of electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time Rainfall Evaporation and ambient temperature ;

[0007] Decoupling modeling steps: Based on the slope angle Calculate the hydraulic gradient I based on the soil infiltration rate. The soil temperature Vegetation root system coefficient The soil permeability factor Kr is calculated based on the target irrigation time t; the original multivariate coupled water balance equation is decoupled into independent one-dimensional water balance equations using the hydraulic gradient I and the soil permeability factor Kr, thus obtaining the irrigation water demand. With irrigation time The analytical relationship;

[0008] State prediction step: based on the soil moisture The soil electrical conductivity The combined concentration of nitrogen, phosphorus, and potassium The soil temperature and the soil pH value as initial state variables With the rainfall amount The evaporation amount and the ambient temperature As a measurable disturbance variable Based on the opening degree of the electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time As a control variable A state-space model is established based on the one-dimensional water balance equation. Model parameters are calculated using the hydraulic gradient I and the soil permeability factor Kr, and the state prediction equation is then used. The state is iteratively predicted, where... , ;

[0009] Rolling optimization steps: With the goal of minimizing the deviation between the predicted value obtained by the iterative prediction and the target value, construct a constrained rolling time-domain objective function J, solve for the optimal control quantity at the current time, and output the execution;

[0010] Feedback correction step: After executing the current control variable, proceed to the next sampling cycle, using the latest actual sensor state output from the data acquisition and preprocessing steps. The predicted value output by the state prediction step Calculate prediction error Using the prediction error The state prediction equation is calibrated online, and the calibrated model is used for rolling optimization in the next time step.

[0011] Hybrid deployment steps: The decoupled modeling step, state prediction step, rolling optimization step, and feedback correction step are deployed on the edge computing platform to run in real time, and the data is uploaded to the cloud. The data-driven model trained in the cloud is used to assist in adjusting the weight parameters in the rolling optimization step.

[0012] Furthermore, the filtering and noise reduction process includes: using a Savitzky-Golay smoothing filter to remove high-frequency noise from the original sensor signal, with a window length of 7 and an order of 3; using a fast Fourier transform to identify and filter out the characteristic frequency bands of periodic interference; the historical window length of the sliding window dataset is 180s, and the prediction window length is 60s.

[0013] Furthermore, the hydraulic gradient The calculation formula is: ,in The slope angle

[0014] Degree; the formula for calculating the soil permeability factor Kr is: ,in To measure the soil infiltration rate, The soil temperature influence coefficient. The root and stem divergence coefficient is denoted by t, where t is the target irrigation time; the soil temperature influence coefficient is also mentioned. The value is: when the soil temperature At ≤5℃, =0.3; when the soil temperature At >5℃, =1+0.05×( -20); the rhizome divergence coefficient The values ​​are: 0.95 for sowing stage, 0.9 for grassland surface irrigation, and 0.8 for grassland drip irrigation.

[0015] Furthermore, the expression for the rolling time-domain objective function J is:

[0016] ;in To predict the time domain, To control the time domain and ≤ , To control the rate of change of quantity, , The weight matrix is ​​positive definite and is determined through offline simulation or reinforcement learning. The positive definite diagonal weight matrix is ​​determined through offline simulation or reinforcement learning; the initial tuning principle of the weight matrix W3 is as follows: ,in Let be the maximum permissible rate of change for the i-th control variable.

[0017] Furthermore, the constraints include at least: valve opening degree: Single irrigation time: Soil moisture: Soil electrical conductivity: μS / cm; Combined nitrogen, phosphorus, and potassium concentrations: mg / kg; Soil temperature: Soil pH value: Valve actuation rate: Water pump safety pressure: .

[0018] Furthermore, the prediction error is utilized in the feedback correction step. The state prediction equation is corrected online using one of the following methods:

[0019] (1) Error coefficient method: The prediction error is calculated using the error coefficient method. The coefficient γ is superimposed into subsequent predictions, and the corrected prediction is: , where 0 < γ < 1, and the typical value γ = 0.2;

[0020] (2) Kalman filtering method: Design a Kalman filter to estimate state deviation and model parameter drift online.

[0021] Furthermore, the edge computing platform is an ARM Cortex-M7 or Raspberry Pi CM4, running an embedded real-time operating system, with a single-step computation time of less than 50ms and a memory footprint of less than 2MB; the data-driven model trained in the cloud includes: using TimeLLM or PATCHTST for short-term time series prediction, using PPO reinforcement learning algorithm for long-term policy optimization, training a control policy network with the goal of minimizing unit irrigation cost, and extracting a rule table or a small-scale neural network to deploy to the edge controller.

[0022] A smart water-saving irrigation system for green slopes based on multivariate decoupling and predictive control, used to execute the method described above, includes:

[0023] The data acquisition module is set in each grid cell of the slope and is used to collect multi-source sensor data;

[0024] The data preprocessing module is used to filter and reduce noise in the collected multi-source sensor data and unify the timestamps, outputting the soil moisture after unifying the timestamps. Soil electrical conductivity Combined concentration of nitrogen, phosphorus and potassium Soil temperature Soil pH value Actual burial depth of the sensor Slope angle Opening degree of electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time Rainfall Evaporation and ambient temperature ;

[0025] The decoupled modeling module is used to determine the slope angle. Calculate the hydraulic gradient I based on the soil infiltration rate. The soil temperature Vegetation root system coefficient The soil permeability factor Kr is calculated based on the target irrigation time t; the original multivariate coupled water balance equation is decoupled into independent one-dimensional water balance equations using the hydraulic gradient I and the soil permeability factor Kr, thus obtaining the irrigation water demand. With irrigation time The analytical relationship;

[0026] State prediction module, used to predict the soil moisture The soil electrical conductivity The combined concentration of nitrogen, phosphorus, and potassium The soil temperature and the soil pH value as initial state variables With the rainfall amount The evaporation amount and the ambient temperature As a measurable disturbance variable Based on the opening degree of the electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time As a control variable A state-space model is established based on the one-dimensional water balance equation. Model parameters are calculated using the hydraulic gradient I and the soil permeability factor Kr, and the state prediction equation is then used. The state is iteratively predicted, where... , It is a time-varying matrix;

[0027] The rolling optimization module is used to construct a constrained rolling time-domain objective function J with the goal of minimizing the deviation between the predicted value obtained by the iterative prediction and the target value, and solve for the optimal control quantity at the current time.

[0028] An actuator, connected to the rolling optimization module, is used to receive and execute the optimal control quantity;

[0029] The feedback correction module is connected to both the data preprocessing module and the state prediction module, and is used to apply the actual sensor state output by the data preprocessing module. The predicted value output by the state prediction module Calculate the prediction error, use the prediction error to correct the state prediction model, and feed the corrected parameters back to the state prediction module;

[0030] The edge computing platform integrates the data preprocessing module, the decoupling modeling module, the state prediction module, the rolling optimization module, and the feedback correction module.

[0031] The cloud optimization module is connected to the edge computing platform via a communication network. It is used to train the data-driven model and send parameters to the edge computing platform to assist in adjusting the weight parameters in the rolling optimization module.

[0032] Furthermore, the data acquisition module includes a multi-parameter sensor group, tilt sensor, electric valve opening sensor, pipeline flow sensor, pipeline pressure sensor, and small weather station deployed in each grid cell; the multi-parameter sensor group includes a soil moisture sensor, soil conductivity sensor, nitrogen, phosphorus and potassium comprehensive concentration sensor, soil temperature sensor, soil pH sensor, and sensor burial depth detection unit; the data preprocessing module includes a Savitzky-Golay smoothing filter unit and a fast Fourier transform filter unit.

[0033] Furthermore, the actuator includes an electric ball valve and a variable frequency water pump installed in each irrigation area; the opening range of the electric ball valve is 0° to 90°, with 0° corresponding to closed and 90° corresponding to fully open; the edge computing platform is an ARM Cortex-M7 or Raspberry Pi CM4, running an embedded real-time operating system and integrating an embedded QP solver; the communication network is a 4G DTU or a 5G communication module; the system also includes a solar power supply module, which includes a photovoltaic panel and a battery, used to provide power to the edge computing platform, the data acquisition module, and the actuator.

[0034] The advantages of this invention are:

[0035] (1) Significantly reduce the complexity of multivariate coupling. By using the analytical formulas for soil permeability factor Kr and hydraulic gradient I, the more than 20 mutually coupled physical quantities (soil texture, sensor burial depth, slope gradient, soil temperature, soil moisture, root characteristics, pipeline water pressure, etc.) in slope irrigation are decoupled into one-dimensional water balance equations for each region. After decoupling, the computational load is reduced by two orders of magnitude, enabling model predictive control (MPC), which originally required high-performance servers to run, to be executed in real time on low-cost edge controllers.

[0036] (2) Effective compensation for large time lags and inter-regional interference. An N-step rolling time-domain predictive control is employed to predict infiltration delays (over 30 minutes) and the impact of slope runoff migration in advance, and the regional coupling interference term is explicitly considered in the objective function. Compared to traditional PID or fuzzy control, this invention avoids upstream waterlogging and downstream drought on the slope, reduces the standard deviation of soil moisture between regions by more than 60%, and eliminates fertilizer loss caused by ineffective runoff.

[0037] (3) Adaptable to various slope conditions and robust. The Kr and I formulas include parameters such as slope gradient, soil temperature, and vegetation type, which can be calibrated for different engineering sites. The model predictive control has built-in constraint processing capabilities and can automatically adapt to changes in soil texture, sensor drift, valve aging, etc., without the need for frequent manual calibration.

[0038] (4) Supports edge computing and hybrid intelligence. The intelligent irrigation algorithm of this invention can run in real time on an ARM embedded platform (single chip cost < 100 yuan) without the need for a high-end server. It also supports hybrid deployment with large models such as TimeLLM and PPO or reinforcement learning, learns global strategies in the cloud, and executes deterministic MPC at the edge, taking into account both low latency and continuous optimization capabilities, making it very suitable for unattended scenarios of plant management on remote green slopes.

[0039] (5) Save water and fertilizer, and protect water and soil resources. Soil moisture, pH, electrical conductivity and other factors are always kept within safe ranges during the control process. By optimizing the irrigation water volume and time of each area in real time, and by automatically skipping ineffective irrigation in combination with rainfall sensors, water can be saved by 20% to 30% compared with conventional timed irrigation; by controlling soil moisture within a suitable range, the utilization rate of nitrogen, phosphorus and potassium fertilizers is increased by more than 15%, and non-point source pollution is reduced. Attached Figure Description

[0040] Figure 1 This is an overall control flowchart of an intelligent water-saving irrigation method for green slopes based on multivariable decoupling and predictive control according to an embodiment of the present invention.

[0041] Figure 2 This is a flowchart of the multi-region coupled irrigation judgment process for slopes according to an embodiment of the present invention;

[0042] Figure 3 This is a flowchart of the decoupling factor modeling process according to an embodiment of the present invention;

[0043] Figure 4 This is a flowchart of the predictive control rolling optimization process according to an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. 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.

[0045] Please see Figure 1 This invention provides an intelligent water-saving irrigation method for green slopes based on multivariate decoupling and predictive control. First, by constructing analytical correction formulas for soil permeability factor Kr and hydraulic gradient I, more than 20 coupled variables in slope irrigation are decoupled into equivalent single-variable water balance equations for each region, reducing computational complexity by two orders of magnitude. Then, based on the decoupled state-space model, rolling time-domain model predictive control (MPC) is used for constrained online optimization to compensate for large time lags and inter-regional disturbances (such as...). Figure 2 (as shown); Finally, it supports lightweight edge computing and reinforcement learning hybrid deployment. The method specifically includes the following steps:

[0046] Step 1: Multi-source data acquisition and preprocessing

[0047] (1) Data Acquisition: The target greening slope irrigation area is divided into N units (numbered n=1,2,…,N from top to bottom of the slope) according to the grid division method. Each unit is equipped with a multi-parameter sensor group to collect multi-source data in real time, including: soil moisture. (Unit: %, range 0~100%); Soil electrical conductivity (Unit: μS / cm, range 0~2000 μS / cm); Combined nitrogen, phosphorus, and potassium concentrations (Unit: mg / kg, range 0~20000 mg / kg); Soil temperature (Unit: °C, range -30~80 °C); Soil pH value (Measuring range 3~10); Actual burial depth of the sensor (Unit: cm, varies depending on soil type); Slope angle (Unit: °, obtained through tilt sensor or design drawings); Opening degree of electric valves in each area (Unit: °, 0° closed, 90° fully open); each area (Unit: Pa) Irrigation water requirement (Unit: m) 3 ), target irrigation time (Unit: min); Rainfall (Ambient temperature can be obtained through small weather stations or by connecting to public weather forecast interfaces) (Unit: °C) Evaporation (Unit: mm)

[0048] (2) Preprocessing: The raw sensor signal was smoothed using a Savitzky-Golay smoothing filter (window length 7, order 3) to remove high-frequency noise; periodic interference (such as pressure pulsation caused by pump start-up and shutdown) was identified and filtered out using Fast Fourier Transform (FFT); a sliding window dataset was constructed with a historical window length of 180s and a prediction window length of 60s to capture the penetration delay effect. All data were stored in the edge controller's memory database after being synchronized.

[0049] Step 2: Multivariable decoupling modeling

[0050] like Figure 3As shown, in response to the multivariate strong coupling problem in slope irrigation, this invention proposes and defines two core decoupling factors—soil permeability factor Kr and hydraulic gradient I. Their physical meanings are as follows: Kr comprehensively represents the soil's ability to allow water to infiltrate vertically (affected by soil type, temperature, plant roots, hydraulic drive, etc.); I represents the ratio of the driving force of water flow along the slope direction to friction loss under slope conditions.

[0051] Formula for calculating hydraulic gradient I (fitted from slope hydraulic tests):

[0052]

[0053] in, The slope is (°). This indicates that gravity is driven by the tangential component along the slope. This is expressed as an empirical friction correction term (1.87% is the average value calibrated for various soil types and roughness). When When the angle is 0° (horizontal ground), I = 0.0187, maintaining the minimum water flow driving force; when When the angle is 45°, I = 0.3667, the driving force increases significantly, which is consistent with engineering practice.

[0054] Formula for calculating soil permeability factor Kr:

[0055]

[0056] in, To measure the soil infiltration velocity (unit: cm / s), it can be obtained through a pre-conducted single-point infiltration test or by referring to a table based on the soil texture; The soil temperature influence coefficient. This reflects the effect of water temperature on viscosity; when ≤5℃ (high-altitude and cold regions) ; The root and stem divergence coefficient is determined based on vegetation type and growth stage: 0.95 for sowing stage, 0.9 for grassland surface irrigation, and 0.8 for grassland drip irrigation. The target irrigation time (in minutes) is given by the initial value of the upper-level predictive control and then iteratively corrected.

[0057] By introducing Kr and I, the original multivariate coupled formula for single-region irrigation water demand can be decoupled into the following simplified form:

[0058]

[0059]

[0060] in, Let be the irrigation water requirement (m³) for the nth region. The target irrigation time (min) for the nth region; is the irrigation area of ​​the nth region (m²); L is the sensor burial depth (cm). The water flow tension correction factor (empirical value is 0.02) In practice, take an empirical value of 0.02 multiplied by the basic tension coefficient or make minor adjustments based on the soil suction curve. Disturbance flow (including lateral inflow / outflow from adjacent areas, rainfall replenishment, evaporation loss, etc.) is dynamically estimated from sensor data, such as... Figure 2 (As shown). This decoupling operation simplifies the original nonlinear equation system, which required solving 20 coupled variables simultaneously, into a one-dimensional water balance calculation independent of each region, reducing the computational complexity from O(N³) to O(N), thus laying the foundation for real-time edge control.

[0061] Step 3: Establishing the State-Space Predictive Control Model

[0062] Based on multivariate decoupling, a discrete-time state-space model is established for each region to predict future changes in soil moisture over a finite time domain.

[0063] Define global state variables and control variables ( (Sampling time, sampling period 30-180s)

[0064] State variables include: ; ; ; ; .

[0065] control variables include: ; ; ; .

[0066] Measurable interference variables include: ; ; .

[0067] The linearized state prediction equation is:

[0068]

[0069] in, , Alternatively, online identification can be performed using recursive least squares.

[0070] For soil moisture, the core control variable, the prediction formula is derived based on the decoupled water balance:

[0071]

[0072] in, This refers to the input parameters for controlling a certain area, specifically the equivalent flow rate when the valve is open.

[0073] After N iterations of prediction:

[0074]

[0075] Step 4: Constrained Scrolling Time-Domain Optimization Solution

[0076] At each sampling time The controller determines the current state. Solve the optimal control problem in the finite time domain.

[0077] Based on minimizing moisture deviation and water resource and energy consumption costs, a rolling time-domain objective function is proposed. :

[0078]

[0079] in, For prediction in the time domain; To control the time domain ( ≤ ); For example, soil moisture. Target range 30%~50%, conductivity Target <1500μS / cm; To control the rate of change of quantity, used to suppress frequent valve operation; , The weight matrix is ​​positive definite and is determined through offline simulation or reinforcement learning. This is a positive definite diagonal weight matrix, where the diagonal elements correspond to the adjustment coefficients for the rate of change of each control variable (valve opening, irrigation time, and pump pressure). A larger value results in smoother actuator movement and a longer valve lifespan, but also a slower response. The smaller the value, the faster the system response, but the more wear and tear on the actuator. The initial tuning principle is based on the maximum allowable operating frequency of the actuator as a constraint. ,in Let be the maximum allowable rate of change for the i-th control variable. Subsequent adjustments will be made using cloud-based reinforcement learning or local adaptive algorithms. , Adjust simultaneously.

[0080] Rolling time domain objective function The constraint condition is: valve opening: Single irrigation time: Soil moisture: Soil electrical conductivity: μS / cm; Combined nitrogen, phosphorus, and potassium concentrations: mg / kg; Soil temperature: Soil pH value: Valve actuation rate: Water pump safety pressure: .

[0081] Rolling time domain objective function It can be transformed into a standard quadratic programming (QP) form and solved in milliseconds using embedded QP solvers (such as OSQP, qpOASES) or fast algorithms based on gradient projection. The optimal control sequence is obtained from the solution. , , However, only the first control variable is executed. (That is, the opening degree of each valve, the opening time, the water pump pressure, etc. at the current moment).

[0082] Step 5: Rolling Time-Domain Feedback and Model Correction

[0083] like Figure 4 As shown, execute After that, the system enters the next sampling cycle. +1). The controller reads the actual sensor status. Contrary to previous predictions Compare and calculate the prediction error:

[0084]

[0085] Model calibration can be performed using one of the following two methods:

[0086] (1) Error coefficient method: The error With coefficient (0< <1, typical value 0.2) are superimposed into subsequent predictions, i.e., the corrected prediction is .

[0087] (2) Kalman filtering method: If the system model has high accuracy, a Kalman filter can be designed to estimate the state deviation and model parameter drift online, thereby further improving robustness.

[0088] The corrected model is used for rolling optimization at the next time step, forming a closed-loop control.

[0089] Step Six: Lightweight Deployment and Intelligent Enhancement

[0090] This invention fully considers the computational resource constraints of slope scenarios and designs a two-level computational architecture:

[0091] (1) Edge level (real-time control): Deployed on low-power platforms such as ARM Cortex-M7 or Raspberry Pi CM4, running the above decoupled calculation and QP solution, with a typical single-step calculation time of <50ms, memory usage of <2MB, and a sampling period of 60s, which can meet the real-time requirements.

[0092] (2) Cloud / Server Level (Non-Real-Time Optimization): For nonlinear disturbances that are difficult to model (such as extreme weather, pipeline blockage) or scenarios that require long-term strategy optimization, a data-driven approach is used for offline training, and then a lightweight model is generated and deployed to the edge level.

[0093] 1) Short-term time series prediction: Using TimeLLM or PATCHTST (with less than 1 billion parameters), after training on RTX 4090, it is exported as an edge-runnable inference model through ONNX for interference prediction in rolling optimization.

[0094] 2) Long-term strategy optimization: Reinforcement learning (PPO algorithm) is adopted with the goal of minimizing the unit irrigation cost. The control policy network is trained in a simulation environment, and then the rule table or a small-scale neural network is extracted into the edge controller to achieve adaptive adjustment of the weight matrix. , , Add adjustment terms to the objective function for the actuator action frequency and the cumulative number of valve actions. This is to balance irrigation effectiveness with equipment lifespan.

[0095] Through the above hybrid architecture, the present invention ensures both the determinism and reliability of real-time control, and also has the ability to continuously learn and evolve.

[0096] This invention also provides an intelligent water-saving irrigation system for green slopes based on multivariable decoupling and predictive control, used to implement the method described above, including:

[0097] The data acquisition module is set in each grid cell of the slope and is used to collect multi-source sensor data;

[0098] The data preprocessing module is used to filter and reduce noise in the collected multi-source sensor data and unify the timestamps, outputting the soil moisture after unifying the timestamps. Soil electrical conductivity Combined concentration of nitrogen, phosphorus and potassium Soil temperature Soil pH value Actual burial depth of the sensor Slope angle Opening degree of electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time Rainfall Evaporation and ambient temperature ;

[0099] The decoupled modeling module is used to determine the slope angle. Calculate the hydraulic gradient I based on the soil infiltration rate. The soil temperature Vegetation root system coefficient The soil permeability factor Kr is calculated based on the target irrigation time t; the original multivariate coupled water balance equation is decoupled into independent one-dimensional water balance equations using the hydraulic gradient I and the soil permeability factor Kr, thus obtaining the irrigation water demand. With irrigation time The analytical relationship;

[0100] State prediction module, used to predict the soil moisture The soil electrical conductivity The combined concentration of nitrogen, phosphorus, and potassium The soil temperature and the soil pH value as initial state variables With the rainfall amount The evaporation amount and the ambient temperature As a measurable disturbance variable Based on the opening degree of the electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time As a control variable A state-space model is established based on the one-dimensional water balance equation. Model parameters are calculated using the hydraulic gradient I and the soil permeability factor Kr, and the state prediction equation is then used. The state is iteratively predicted, where... , It is a time-varying matrix;

[0101] The rolling optimization module is used to construct a constrained rolling time-domain objective function J with the goal of minimizing the deviation between the predicted value obtained by the iterative prediction and the target value, and solve for the optimal control quantity at the current time.

[0102] An actuator, connected to the rolling optimization module, is used to receive and execute the optimal control quantity;

[0103] The feedback correction module is connected to both the data preprocessing module and the state prediction module, and is used to apply the actual sensor state output by the data preprocessing module. The predicted value output by the state prediction module Calculate the prediction error, use the prediction error to correct the state prediction model, and feed the corrected parameters back to the state prediction module;

[0104] The edge computing platform integrates the data preprocessing module, the decoupling modeling module, the state prediction module, the rolling optimization module, and the feedback correction module.

[0105] The cloud optimization module is connected to the edge computing platform via a communication network. It is used to train the data-driven model and send parameters to the edge computing platform to assist in adjusting the weight parameters in the rolling optimization module.

[0106] The following three examples of irrigation projects will be used to illustrate the specifics.

[0107] Example 1: Irrigation Project for Green Slopes of a Highway in Southern China

[0108] Site conditions: The slope is 120m long and 25m wide, with an average slope of 18°, divided into 30 irrigation zones. The soil type is red soil (clay soil), and the sensor is buried at a depth of 12cm. The slope vegetation is Bermuda grass (root depth approximately 15cm). The irrigation target is to maintain soil moisture at 25%~45% and electrical conductivity <1200μS / cm. There is a 50m³ water storage tank at the toe of the slope, with a water pump head of 45m and equipped with a frequency converter. The slope irrigation pipeline actuators are electric ball valves (0~90°, adjustment time 2s), one per zone. The controller is an STM32H743 (Cortex-M7, 400MHz), running the FreeRTOS real-time system, with an integrated OSQP embedded solver. Daily data is uploaded to the central server via a 4G DTU.

[0109] Parameter initialization: Considering the slow permeability of clayey soil, the sampling period is set to 90 seconds. Prediction time domain N=8, control time domain M=3; weight matrix is ​​set as follows: The weight of moisture on the diagonal is 10, and the weight of other states is 1. Valve opening weight 0.5, irrigation time weight 0.1. The weight of the valve opening change rate is 0.5; the valve opening change rate is ≤5° / s, and the maximum irrigation time per cycle is 25 minutes; measured through on-site double-ring infiltration tests. =0.008cm / s, soil temperature coefficient Automatic calculation; root and stem divergence coefficient =0.9 (grassland irrigation).

[0110] Implementation process: After the system is powered on, it automatically reads the initial values ​​of the sensors in each area and calculates the current Kr (the initial value is based on experience, t=10min); a control loop is triggered every 90 seconds: data acquisition, Savitzky-Golay filtering, updating Kr and I, constructing a state-space model, calling OSQP to solve QP, outputting valve opening degree and opening time, and execution; in the first 90-second cycle, it is found that the soil moisture in area 1-1 (first row and first column of the slope top) is 22% (below the target lower limit of 25%), and in area 2-1 (second row and first column of the slope top)... The humidity in the first row (first column) was 28% (normal), but the prediction model considered the downward migration of water, so the valve opening in area 1-1 was only set to 40° (instead of 90°), and the irrigation time was 8 minutes to prevent water from flowing rapidly to area 2-1 and causing overwatering. After 2 hours of actual operation, the soil humidity in area 1-1 rose to 28%, and the soil humidity in area 2-1 rose from 28% to 32%, still within the target range. In the 6th hour, a moderate rain (8 mm / h) suddenly fell, triggering the rainfall sensor. The controller automatically paused the irrigation task and adjusted the settings in the prediction model. The value is set to a positive value and returns to normal 1 hour after the rain stops; at 2 AM every day, the edge controller packages and uploads 24-hour data (sensor timing, control variables, water consumption) to the cloud server; the server runs PPO reinforcement learning every 7 days to fine-tune the weight matrix. , , Generate new parameters and distribute them.

[0111] Results: After 60 days of continuous operation, compared with the timed irrigation of adjacent slopes, the average water saving rate was 26.4%; the standard deviation of soil moisture decreased from 17.2% in the open loop to 5.8%; fertilizer usage was reduced by 22%, while the vegetation coverage rate showed no significant difference; the controller's average solution time was 38ms, with a maximum of 52ms, meeting the 90s cycle requirement.

[0112] Example 2: Irrigation Project for Greening Steep Slopes in an Open-Pit Mine in Northern China

[0113] Site conditions: The slope is 80m high and 60m wide, with an average slope of 52°, divided into 64 irrigation zones. The soil is a mixture of gravelly soil and topsoil, with rapid infiltration; the sensor is buried 20cm deep. The vegetation consists of Amorpha fruticosa and tall fescue, with a root depth of approximately 25cm. The irrigation objective is to maintain soil moisture at 20%–35%, electrical conductivity <1500μS / cm, and to prevent shallow landslides. The water source is a rainwater harvesting pond at the foot of the mountain, with an effective volume of 200m³, a pump head of 90m, high pump power, and a frequency converter. The slope irrigation pipeline actuator is an electric ball valve (0–90°, adjustment time 3s), one per zone. Each zone is equipped with a TEROS 12 three-in-one sensor. The controller is a Raspberry Pi CM4 with real-time patching, an external 4G communication module, and runs embedded Linux and the OSQP solver.

[0114] Parameter initialization: Considering the soil is a mixture of gravelly soil and imported soil, with rapid infiltration, and a sampling period of 40 seconds. Prediction time domain N=12, control time domain M=4; weight matrix set to... On the diagonal line, moisture has a weight of 15, and electrical conductivity has a weight of 5. Valve opening weight 0.3, irrigation time weight 0.05. The weight of the valve opening change rate is 0.8; the valve opening change rate is ≤8° / s, and the maximum irrigation time per cycle is 5 minutes; this was measured through a field double-ring infiltration test. =0.035cm / s, temperature coefficient Automatic calculation; root and stem divergence coefficient =0.85 (grassland irrigation).

[0115] Implementation Process: TEROS 12 sensors were installed in each of the 64 designated areas, along with four main pipelines. The controller was installed in the distribution box at the foot of the slope and communicated with the valves and frequency converter via an RS485 bus. Upon initial power-up, the system performed a self-test to establish the actuator dynamic characteristic library. During the first irrigation cycle, sensor readings revealed that the humidity at the top of the slope was generally 18% (below the lower limit of 20%), 22%–26% in the middle of the slope, and 28%–31% at the bottom. MPC predicted that if the electric ball valve at the top of the slope opened at 60° for 5 minutes of irrigation, the water would flow to the middle of the slope within 40 seconds, exceeding the 35% upper limit. Therefore, the electric ball valve at the top of the slope was opened only 25° for 1.5 minutes of irrigation, while the middle and bottom slopes were closed. After 40 seconds, feedback indicated that the soil moisture at the top of the slope had risen to 20.5%, as expected, but the slope… The soil moisture in the third row slightly exceeded the prediction (26%), rising from 23% to 28%. The calculation error was e = +2%. Kalman filtering was used for correction, and the lateral diffusion coefficient was increased from 0.15 to 0.18. Subsequently, small-volume irrigation was applied to the slope crest every 80 seconds for 1-2 minutes over the next 2 hours. Irrigation was only initiated when the moisture content in the middle of the slope was below 22%, and the opening angle was limited to within 15°. The moisture content in the area of ​​the third row, fourth column suddenly increased to 34%. Investigation revealed that water from adjacent irrigation areas diffused laterally through surface fissures. MPC reduced the valve opening in the third row, third column from 20° to 10° and... The inflow compensation for the fourth column was increased by 0.5 L / min; on the fifth day of operation, a heavy rainfall of 30 mm / h occurred, and the controller immediately stopped irrigation, closed the valve and reduced the pressure to 0.2 MPa for standby. Two hours after the rain stopped, the humidity was still generally above 35%, and it continued to standby until it naturally dropped below 32% to recover; data was uploaded to the server every 14 days, and after the first update of the PPO algorithm, the upper limit of moisture was adjusted from 35% to 38%.

[0116] Results: After 90 days of continuous operation, the average water saving rate was 31.2%, the turbidity of slope runoff decreased by 70%, the standard deviation of soil moisture decreased from 14.3% to 4.1%, no landslides occurred, the average solution time of the controller was 52ms, the vegetation coverage rate reached 92%, and the survival rate of Amorpha fruticosa increased by 18%.

[0117] Example 3: Irrigation Project for Green Slopes along Highways on the Qinghai-Tibet Plateau

[0118] Site conditions: Altitude 4200m, average annual temperature -2℃, slope length 60m, width 15m, slope 28°, soil is glacial till (gravel >50%), divided into 16 4×4 zones, sensor burial depth 15cm. Vegetation is *Leymus chinensis*. Irrigation target: maintain soil moisture 25%~45%, conductivity <2000μS / cm, irrigation prohibited when soil temperature <0℃. Water source is a mountain stream pump station with a head of 50m. Slope irrigation pipeline actuators are electric ball valves (0~90°, adjustment time 4s), one per zone. Controller is MSP432P401R, solar powered (200W photovoltaic + 200Ah battery), with frost protection cable. Daily data is uploaded to the central server via 4G DTU.

[0119] Parameter initialization: Considering the slow permeability of glacial till, the sampling period is 120s, the prediction time domain N=6, and the control time domain M=2; the weight matrix is ​​set as follows: On the diagonal, moisture has a weight of 8, and temperature deviation has a weight of 3. Valve opening weight 0.3, irrigation time weight 0.1. The weight of the valve opening change rate is 0.3; the valve opening change rate is ≤3° / s; the maximum irrigation time per cycle is 30 min; the pump pressure is 0.3~0.6 MPa; the hydraulic gradient I = 0.431; and the temperature coefficient... Piecewise function: ≤5℃ =0.3, >5℃ =1+0.05×( -20), <0℃ =0 and the valve is forcibly closed; the sensor heating strategy is to energize the heating cable for 1 second (0.5W) before sampling. At positive temperature. =0.012cm / s, close to 0 at sub-zero temperatures. =0.92.

[0120] Implementation process: The system was installed in early May after the local soil thawed. An initial 24-hour idle test was performed upon power-up, revealing that the daytime temperature of the glacial till soil was 10°C. The speed can reach 0.012 cm / s. However, at night at -3°C, the sensor freezes and readings become abnormal. Therefore, temperature judgment logic has been added to the software. Skip the area when the temperature is below 0℃ and mark it as "frozen"; wake up at 8:00 AM daily during the frost-free period (deep sleep at night), collect data from 16 areas to calculate Kr, substitute it into the water balance equation, and add evaporation to the interference term. In mid-June, the soil moisture in region 1-1 was 22% (below the lower limit of 25%). MPC predicted that if the electric ball valve was opened to 45° and irrigation lasted for 20 minutes, water would flow to region 2-1 (currently 28%), causing it to exceed 45%. Therefore, the output was: region 1-1 opened to 30° and irrigation lasted for 12 minutes, while region 2-1 was closed. After 120 seconds, the moisture in region 1-1 rose to 26% and region 2-1 rose to 33%, both within safe ranges. In late October, the nighttime temperature suddenly dropped to -4℃. The sensor... When the temperature drops to -2℃, the controller immediately implements "freeze protection," shutting off valves, cutting off power to the water pump, and switching to sleep mode with only soil temperature monitoring enabled. The temperature rises to 3℃ the following morning, but there is a 2-hour delay before resuming operation to allow the ice crystals to melt. The heating cable only operates for 1 second before sampling, with a total heating energy consumption of approximately 360J for 720 daily samplings. Including the MCU power consumption, the daily energy consumption is approximately 10Wh. Photovoltaic power is sufficient on sunny days, and automatically reduces the frequency to 300 seconds and shuts down data reporting on cloudy days. When the humidity deviation exceeds 5% every 7 days, the moisture weight is adjusted by ±0.5. After one month of operation, the weight converges from 8 to 9.2.

[0121] Results: After 4 months of continuous operation during the frost-free period, the average water saving rate was 24.3%, the seedling emergence rate was 82% (compared to 73% in the control group), and the vegetation height and coverage were better than the control group. There was no freeze-thaw damage caused by irrigation (compared to a 15% seedling loss in the control group). The average solution time of the controller was 28ms. Under continuous 7 days of cloudy and rainy conditions, the minimum remaining battery power was 42%, and there was no power outage.

[0122] This invention has the following features and effects:

[0123] Significantly reduced computational complexity: By constructing analytical formulas for soil permeability factor Kr and hydraulic gradient I, more than 20 coupled variables in slope irrigation are decoupled into independent one-dimensional water balance equations for each region. The computational load is reduced from O(N³) to O(N), enabling model predictive control to run in real time on a low-cost edge controller (ARM Cortex-M7 / Raspberry Pi CM4), with single-step calculation <50ms and memory <2MB.

[0124] Effective compensation for large time lags and slope runoff disturbances: Rolling time-domain predictive control (MPC) is adopted to predict infiltration delay (≥30 min) and gravity migration of water flow in advance. The objective function explicitly constrains soil moisture, electrical conductivity, and other indicators in each region. In actual engineering, the standard deviation of soil moisture between regions is reduced by more than 60%, preventing fertilizer loss caused by upstream flooding, downstream drought, and ineffective runoff.

[0125] Adaptive to various slope conditions: The Kr and I formulas include parameters such as slope, soil temperature, and vegetation root system, and can be flexibly calibrated for different soil types (clay, gravelly soil, glacial till) and climates (humid and hot in the south, steep in the north, and frozen soil in the plateau), with strong robustness.

[0126] Edge-cloud hybrid intelligence: The edge performs deterministic MPC, while the cloud uses TimeLLM / PATCHTST for short-term disturbance prediction and PPO reinforcement learning for long-term optimization of the weight matrix, balancing real-time performance and continuous evolution capabilities, making it suitable for remote unattended scenarios.

[0127] Significant water and fertilizer savings: By combining rainfall / evaporation compensation, ineffective irrigation is avoided; soil moisture is controlled within a suitable range, increasing nitrogen, phosphorus, and potassium utilization rates by more than 15%. Engineering verification shows water savings of 24%–31%, fertilizer usage reduced by 22%, and vegetation coverage significantly increased, while reducing non-point source pollution and soil erosion.

[0128] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A smart water-saving irrigation method for green slopes based on multivariate decoupling and predictive control, characterized in that, include: Data acquisition and preprocessing steps: Divide the slope into N units, collect multi-source sensor data from each unit and perform filtering and noise reduction, construct a sliding window dataset and unify the timestamps to obtain soil moisture after unifying the timestamps. Soil electrical conductivity Combined concentration of nitrogen, phosphorus and potassium Soil temperature Soil pH value Actual burial depth of the sensor Slope angle Opening degree of electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time Rainfall Evaporation and ambient temperature ; Decoupling modeling steps: Based on the slope angle Calculate the hydraulic gradient I based on the soil infiltration rate. The soil temperature Vegetation root system coefficient The soil permeability factor Kr is calculated based on the target irrigation time t; using the hydraulic gradient I and the soil permeability factor Kr, the original multivariate coupled water balance equation is decoupled into independent one-dimensional water balance equations for each unit, thus obtaining the irrigation water demand. With irrigation time The analytical relationship; State prediction step: based on the soil moisture The soil electrical conductivity The combined concentration of nitrogen, phosphorus, and potassium The soil temperature and the soil pH value As initial state variables With the rainfall amount The evaporation amount and the ambient temperature As a measurable disturbance variable Based on the opening degree of the electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time As a control variable A state-space model is established based on the one-dimensional water balance equation. Model parameters are calculated using the hydraulic gradient I and the soil permeability factor Kr, and the state prediction equation is then used. The state is iteratively predicted, where... , ; Rolling optimization steps: With the goal of minimizing the deviation between the predicted value obtained by the iterative prediction and the target value, construct a constrained rolling time-domain objective function J, solve for the optimal control quantity at the current time, and output the execution; Feedback correction step: After executing the current control variable, proceed to the next sampling cycle, using the latest actual sensor state output from the data acquisition and preprocessing steps. The predicted value output by the state prediction step Calculate prediction error Using the prediction error The state prediction equation is calibrated online, and the calibrated model is used for rolling optimization in the next time step. Hybrid deployment steps: The decoupled modeling step, state prediction step, rolling optimization step, and feedback correction step are deployed on the edge computing platform to run in real time, and the data is uploaded to the cloud. The data-driven model trained in the cloud is used to assist in adjusting the weight parameters in the rolling optimization step.

2. The intelligent water-saving irrigation method for green slopes based on multivariable decoupling and predictive control according to claim 1, characterized in that, The filtering and noise reduction process includes: using a Savitzky-Golay smoothing filter to remove high-frequency noise from the original sensor signal, with a window length of 7 and an order of 3; using a fast Fourier transform to identify and filter out the characteristic frequency bands of periodic interference; the historical window length of the sliding window dataset is 180s, and the prediction window length is 60s.

3. The intelligent water-saving irrigation method for green slopes based on multivariate decoupling and predictive control according to claim 1, characterized in that, The hydraulic gradient The calculation formula is: ,in The slope angle is given; the soil permeability factor Kr is calculated using the following formula: ,in To measure the soil infiltration rate, The soil temperature influence coefficient. The root and stem divergence coefficient is denoted by t, where t is the target irrigation time; the soil temperature influence coefficient is also mentioned. The value is: when the soil temperature At ≤5℃, =0.3; when the soil temperature At >5℃, =1+0.05×( -20); the rhizome divergence coefficient The values ​​are: 0.95 for sowing stage, 0.9 for grassland surface irrigation, and 0.8 for grassland drip irrigation.

4. The intelligent water-saving irrigation method for green slopes based on multivariate decoupling and predictive control according to claim 1, characterized in that, The expression for the rolling time-domain objective function J is: ,in To predict the time domain, To control the time domain and ≤ , To control the rate of change of quantity, , The weight matrix is ​​positive definite and is determined through offline simulation or reinforcement learning. The positive definite diagonal weight matrix is ​​determined through offline simulation or reinforcement learning; the initial tuning principle of the weight matrix W3 is as follows: ,in Let be the maximum permissible rate of change for the i-th control variable.

5. The intelligent water-saving irrigation method for green slopes based on multivariate decoupling and predictive control according to claim 4, characterized in that, The constraints include at least: valve opening degree. Single irrigation time: Soil moisture: Soil electrical conductivity: μS / cm; Combined nitrogen, phosphorus, and potassium concentrations: mg / kg; Soil temperature: Soil pH value: Valve actuation rate: Water pump safety pressure: .

6. The intelligent water-saving irrigation method for green slopes based on multivariate decoupling and predictive control according to claim 1, characterized in that, The feedback correction step utilizes the prediction error. The state prediction equation is corrected online using one of the following methods: (1) Error coefficient method: The prediction error is calculated using the error coefficient method. The coefficient γ is superimposed into subsequent predictions, and the corrected prediction is: , where 0 < γ < 1, and the typical value γ = 0.2; (2) Kalman filtering method: Design a Kalman filter to estimate state deviation and model parameter drift online.

7. The intelligent water-saving irrigation method for green slopes based on multivariable decoupling and predictive control according to claim 1, characterized in that, The edge computing platform is an ARM Cortex-M7 or Raspberry Pi CM4, running an embedded real-time operating system, with a single-step computation time of less than 50ms and a memory footprint of less than 2MB; the cloud-trained data-driven model includes: using TimeLLM or PATCHTST for short-term time series prediction, using PPO reinforcement learning algorithm for long-term policy optimization, training a control policy network with the goal of minimizing unit irrigation cost, and extracting rule tables or small-scale neural networks for deployment to the edge controller.

8. A smart water-saving irrigation system for green slopes based on multivariable decoupling and predictive control, used to execute the method according to any one of claims 1-7, characterized in that, include: The data acquisition module is set in each grid cell of the slope and is used to collect multi-source sensor data; The data preprocessing module is used to filter and reduce noise in the collected multi-source sensor data and unify the timestamps, outputting the soil moisture after unifying the timestamps. Soil electrical conductivity Combined concentration of nitrogen, phosphorus and potassium Soil temperature Soil pH value Actual burial depth of the sensor Slope angle Opening degree of electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time Rainfall Evaporation and ambient temperature ; The decoupled modeling module is used to determine the slope angle. Calculate the hydraulic gradient I based on the soil infiltration rate. The soil temperature Vegetation root system coefficient The soil permeability factor Kr is calculated based on the target irrigation time t; using the hydraulic gradient I and the soil permeability factor Kr, the original multivariate coupled water balance equation is decoupled into independent one-dimensional water balance equations for each unit, thus obtaining the irrigation water demand. With irrigation time The analytical relationship; State prediction module, used to predict the soil moisture The soil electrical conductivity The combined concentration of nitrogen, phosphorus, and potassium The soil temperature and the soil pH value As initial state variables With the rainfall amount The evaporation amount and the ambient temperature As a measurable disturbance variable Based on the opening degree of the electric valves in each area Main pipeline water flow pressure Irrigation water demand Target irrigation time As a control variable A state-space model is established based on the one-dimensional water balance equation. Model parameters are calculated using the hydraulic gradient I and the soil permeability factor Kr, and the state prediction equation is then used. The state is iteratively predicted, where... , It is a time-varying matrix; The rolling optimization module is used to construct a constrained rolling time-domain objective function J with the goal of minimizing the deviation between the predicted value obtained by the iterative prediction and the target value, and solve for the optimal control quantity at the current time. An actuator, connected to the rolling optimization module, is used to receive and execute the optimal control quantity; The feedback correction module is connected to both the data preprocessing module and the state prediction module, and is used to apply the actual sensor state output by the data preprocessing module. The predicted value output by the state prediction module Calculate the prediction error, use the prediction error to correct the state prediction model, and feed the corrected parameters back to the state prediction module; The edge computing platform integrates the data preprocessing module, the decoupling modeling module, the state prediction module, the rolling optimization module, and the feedback correction module. The cloud optimization module is connected to the edge computing platform via a communication network. It is used to train the data-driven model and send parameters to the edge computing platform to assist in adjusting the weight parameters in the rolling optimization module.

9. The intelligent water-saving irrigation system for green slopes based on multivariable decoupling and predictive control according to claim 8, characterized in that, The data acquisition module includes a multi-parameter sensor group, tilt sensor, electric valve opening sensor, pipeline flow sensor, pipeline pressure sensor, and small weather station deployed in each grid cell; the multi-parameter sensor group includes a soil moisture sensor, soil conductivity sensor, nitrogen, phosphorus and potassium comprehensive concentration sensor, soil temperature sensor, soil pH sensor, and sensor burial depth detection unit; the data preprocessing module includes a Savitzky-Golay smoothing filter unit and a fast Fourier transform filter unit.

10. The intelligent water-saving irrigation system for green slopes based on multivariable decoupling and predictive control according to claim 8, characterized in that, The actuator includes an electric ball valve and a variable frequency water pump installed in each irrigation area; the opening range of the electric ball valve is 0° to 90°, with 0° corresponding to closed and 90° corresponding to fully open; The edge computing platform is an ARM Cortex-M7 or Raspberry Pi CM4, running an embedded real-time operating system and integrating an embedded QP solver; the communication network is a 4G DTU or a 5G communication module; the system also includes a solar power module, which includes a photovoltaic panel and a battery, for providing power to the edge computing platform, the data acquisition module and the actuator.

Images