Agricultural irrigation control method and system

Abstract

The present invention discloses an agricultural irrigation control method and system, which relate to the field of irrigation control technology. The method and system comprise the following steps: constructing a crop water demand model based on the fusion of crop physiological sensor data and a data-driven model to output a target irrigation flow; calculating the control error between the target irrigation flow and the actual collected flow, and generating a control signal through a model prediction control strategy; inputting a static mapping model fitted by a Sigmoid function, superimposing the output of a dynamic correction model driven by state perception data, and generating a final driving signal; collecting feedback data, performing a closed-loop comparison between the driving signal and the feedback data, and adaptively optimizing irrigation control strategy parameters through a reinforcement learning controller; the present invention dynamically constructs a water demand model that integrates spatial and temporal characteristics by fusing crop physiological sensor data with the data-driven model, and combines a static and dynamic composite control mapping mechanism to solve the problems of agricultural irrigation that cannot be precisely controlled and irrigation strategy rigidity.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural irrigation control, and more particularly, to an agricultural irrigation control method and system. Background Art

[0002] In modern agricultural production, water resources are a key constraint on crop yields and sustainable agricultural development. Their efficient utilization has become a core issue in global agricultural technology development. Particularly in arid and semi-arid regions, traditional irrigation methods such as flooding and furrow irrigation suffer from extensive water allocation, significant water waste, and a high risk of secondary soil salinization, making them unable to meet the precise water control requirements of modern precision agriculture. With the development of cutting-edge information technologies such as the Internet of Things, sensor technology, communication technology, and artificial intelligence, intelligent irrigation systems integrate sensors, communication networks, automatic control equipment, and intelligent decision-making algorithms to automate, refine, and intelligently manage irrigation processes for farmland, green spaces, horticultural crops, and other areas. Their core goal is to dynamically adjust irrigation volume and timing based on information such as crop water requirements, soil moisture, and weather forecasts, thereby conserving water, increasing crop yields, and boosting resource efficiency.

[0003] For example, the wireless intelligent irrigation system disclosed in patent publication number CN102499028B includes a server host, a central controller, wireless electromagnetic valves, and a data acquisition system. The server host encodes the irrigation signal and transmits it to the central controller via a GSM module. Upon receiving the signal, the central controller sends a handshake signal to the server host, processes the received command, decodes it, and transmits it via a wireless transmission module to the wireless electromagnetic valves, controlling their opening for irrigation. The system also includes a data acquisition system for collecting various environmental information. This system selectively opens or closes conventional irrigation electromagnetic valves by directly sending digital control signals. This allows for more targeted irrigation based on plant water requirements, avoiding the waste of precious water resources. Furthermore, the central controller can be controlled by wirelessly editing program commands from the server host or a mobile terminal, such as a mobile phone, eliminating the increased labor and material costs associated with laying control cables, significantly reducing costs.

[0004] For example, the invention patent with announcement number CN119717545A discloses a solenoid valve control system and method for agricultural irrigation. The method determines the hysteresis tolerance point when the solenoid valve controlling the irrigation sprinkler generates control hysteresis based on the solenoid valve's real-time information and control status data; coordinates and corrects the solenoid valve's action time based on the action limit interval and the irrigation spray volume, obtaining a coordinated correction feature when the solenoid valve controlling the irrigation sprinkler generates control hysteresis; and determines the control fitting threshold when the solenoid valve controls irrigation spray at different action times based on the coordinated correction feature; and performs fault-tolerant control of the solenoid valve controlling the irrigation sprinkler based on the action feedback deviation. This application can monitor the solenoid valve's action feedback in real time during control, identify and correct control hysteresis, perform precise coordinated corrections, and dynamically adjust the solenoid valve's control parameters to improve the control stability of the irrigation system.

[0005] The above disclosed technical solutions have at least the following technical problems:

[0006] In current intelligent irrigation systems, proportional solenoid valves, as key actuators, exhibit significant nonlinear and time-varying dynamic characteristics between their control signals and actual drive responses. Traditional linear control methods struggle to guarantee control accuracy, and environmental changes affect solenoid valve performance, further complicating control. This makes it difficult for intelligent irrigation systems to precisely control water supply based on actual crop water requirements, leading to over- and under-watering. Furthermore, irrigation strategies lack feedback-based adaptive online optimization capabilities, making them unable to adapt to changes in the environment and equipment status, hindering precise irrigation control. Existing control frameworks rely heavily on empirical rules, lacking efficient model predictive control that integrates spatial-temporal characteristics and physiological information, making it difficult to achieve a balance between water conservation and precision irrigation.

[0007] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0008] To overcome the above-mentioned shortcomings of the prior art, an embodiment of the present invention provides an agricultural irrigation control method and system. By integrating crop physiological sensor data with a data-driven model, a water demand model that integrates spatial and temporal characteristics is dynamically constructed. Combined with a static and dynamic composite control mapping mechanism, the method and system solve the problems of agricultural irrigation that cannot be precisely controlled and the rigidity of irrigation strategies.

[0009] In the first aspect, the present application provides an agricultural irrigation control method, comprising the following steps: constructing a crop water requirement model based on the fusion of crop physiological sensor data and a data-driven model, and outputting a target irrigation flow in real time; calculating the control error between the target irrigation flow and the actual collected flow, and generating an original control signal of a proportional solenoid valve through a model predictive control strategy; inputting the original control signal into a static mapping model fitted by a Sigmoid function, superimposing the output of a dynamic correction model driven by state perception data, and generating a final drive signal of the proportional solenoid valve; collecting valve execution feedback data, performing a closed-loop comparison between the drive signal and the feedback data, and adaptively optimizing the irrigation control strategy parameters through a reinforcement learning controller.

[0010] In a preferred embodiment, the crop water requirement model is constructed based on the fusion of crop physiological sensor data and a data-driven model to output the target irrigation flow in real time, specifically by collecting meteorological data, soil moisture data, and crop physiological indicator data of the target crop area to form a crop multi-source dataset;

[0011] Multi-source crop data is input into a time series prediction network that integrates graph convolution and Transformer to generate a basic prediction value of crop water requirement. Based on real-time stem flow sensor and transpiration meter data, the basic prediction value is corrected through an adaptive weighting mechanism to output the final value of the crop water requirement model. The target irrigation flow is generated based on the final value.

[0012] In a preferred embodiment, the multi-source crop data is input into the time series prediction network that integrates graph convolution and Transformer to generate a basic prediction value of crop water requirement. Specifically, the target crop area is defined as a node in a topological graph to construct a crop spatial relationship network; the spatial feature embedding vector of the node is extracted through a graph convolution neural network; the multi-source time series data of each crop area is obtained and input into the Transformer time series modeling module, and the time series features are extracted using the self-attention mechanism; the spatial feature embedding vector and the time series features are integrated to generate a basic prediction value of crop water requirement.

[0013] In a preferred embodiment, based on real-time stem flow sensor and transpiration meter data, the basic prediction value is corrected through an adaptive weighting mechanism to output the final value of crop water requirement, specifically: physiological indicator data is obtained, and an equivalent physiological feedback value is generated after normalization processing; the difference between the equivalent physiological feedback value and the basic prediction value of crop water requirement is calculated to generate a crop water stress residual data sequence; a regression correction subnetwork is constructed based on the residual data sequence, and a correction term is output through a gated weighted fusion mechanism; the correction term is weightedly fused with the basic prediction value of crop water requirement to generate a crop water requirement model output value.

[0014] In a preferred embodiment, the control error between the target irrigation flow and the actual collected flow is calculated, and the original control signal of the proportional solenoid valve is generated through the model predictive control strategy, specifically: the current actual irrigation flow of the irrigation system is collected in real time as the current system output; the control error between the target irrigation flow and the actual irrigation flow is calculated; the historical irrigation system control input signal, flow response and external disturbance information are obtained to establish a dynamic prediction model of the irrigation system; based on the dynamic prediction model and the control error, the objective function is constructed and the system constraints are set; the rolling horizon optimization method is used to solve the objective function and generate the optimal control sequence in the future time domain; and the first instruction of the optimal control sequence is extracted as the control signal of the proportional solenoid valve.

[0015] In a preferred embodiment, the original control signal is input into a static mapping model fitted by a Sigmoid function, and the output of a dynamic correction model driven by state perception data is superimposed to generate the final drive signal of the proportional solenoid valve. Specifically, the following steps are performed: historical drive data is obtained, a nonlinear mapping relationship between the original control signal and the drive signal is constructed, and an initial static mapping model is generated based on Sigmoid saturation function fitting; key state variables under the working environment of the solenoid valve are collected in real time, and a set of state perception variables is constructed;

[0016] A dynamic correction model is constructed through a recurrent neural network based on a set of state-sensing variables and first data, wherein the first data includes a set of state-sensing variables, a control signal, and a drive response; an initial static mapping model is used as a feedforward module to output a fitted drive signal, and a dynamic response offset is output through a dynamic correction model; the dynamic response offset is superimposed on the fitted drive signal to generate a final drive response output; and the dynamic model is updated online and corrected in real time using a preset dynamic threshold.

[0017] In a preferred embodiment, the model is updated and dynamically corrected online based on a dynamic threshold, specifically as follows: second data is collected to calculate the driving response error at the current moment, the second data including the real-time control signal, the model output and the actual driving response; an error sliding window is constructed, and whether to trigger the online update is determined based on the error mean and the change trend; after the online update is triggered, a structure that supports online fine-tuning is introduced into the initial static mapping model and the dynamic correction model, the structure including an initial static mapping model using a multi-layer perceptron structure that can be incrementally updated, and a dynamic correction model using a recurrent neural network that supports local parameter fine-tuning; the model training data within the current time window is extracted to construct a training set and a validation set, the model training data including the control signal, the state perception variable sequence and the actual driving response data; a finite-step gradient update of the model parameters is performed based on the training set through a preset micro-batch training strategy; after the update is completed, the model errors before and after the update are compared based on the validation data set, and if the error reduction ratio after the update exceeds a preset ratio threshold, the new model is retained; otherwise, the model parameters are rolled back to the previous optimal model parameters.

[0018] In a preferred embodiment, valve execution feedback data is collected, a closed-loop comparison is performed between the drive signal and the feedback data, and irrigation control strategy parameters are adaptively optimized using a reinforcement learning controller. Specifically, during the irrigation process after the proportional solenoid valve executes the drive signal, multi-dimensional feedback data directly related to the irrigation effect is collected in real time; a deep reinforcement learning strategy optimization network is constructed to generate a reinforcement learning controller.

[0019] According to the output of the reinforcement learning controller, the control strategy parameters of the proportional solenoid valve are dynamically optimized, and the control strategy parameters include the control amplitude of the proportional solenoid valve and the actual irrigation execution duration.

[0020] In a preferred embodiment, the reinforcement learning controller specifically comprises: constructing a state space composed of multidimensional state vectors and an action space composed of control strategy parameters; constructing a reward function for the target irrigation effect through multidimensional feedback data, wherein the reward function includes physiological response data, soil moisture change data, irrigation water use efficiency, and irrigation energy consumption penalty; constructing a reinforcement learning controller based on the state space, action space, and reward function; and obtaining the optimal control strategy parameters by outputting the action that maximizes the reward function in the current state through the reinforcement learning controller.

[0021] In a second aspect, the present application provides a system for an agricultural irrigation control method, comprising: a target irrigation flow acquisition module, a control signal generation module, a control signal conversion module, and a control strategy feedback optimization module, wherein the modules are connected:

[0022] Among them, the target irrigation flow acquisition module is used to build a crop water requirement model based on the fusion of crop physiological sensor data and data-driven models, and output the target irrigation flow in real time; the control signal generation module is used to calculate the control error between the target irrigation flow and the actual collected flow, and generate the original control signal of the proportional solenoid valve through the model prediction control strategy; the control signal conversion module is used to input the original control signal into the static mapping model fitted by the Sigmoid function, superimpose the dynamic correction model output driven by state perception data, and generate the final drive signal of the proportional solenoid valve; the control strategy feedback optimization module is used to collect valve execution feedback data, compare the drive signal with the feedback data in a closed loop, and adaptively optimize the irrigation control strategy parameters through the reinforcement learning controller.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0024] 1. By integrating the graph neural network's ability to model farmland spatial relationships with the Transformer model's strong ability to represent time series dynamics, a crop water requirement prediction model was constructed that has a deep understanding of and precise fitting capabilities for heterogeneous spatiotemporal information. At the same time, by combining the crop physiological sensor feedback mechanism with the target water requirement estimation results, the solenoid valve opening is adaptively adjusted through model predictive control (MPC), achieving dynamic optimization and precise scheduling of irrigation control, significantly improving the irrigation system's intelligent response capability, water resource utilization efficiency, and crop health maintenance level.

[0025] 2. By combining a static nonlinear mapping model based on the Sigmoid function with a dynamic correction model driven by state perception, a high-precision, dynamic adaptive conversion of the proportional solenoid valve control signal to the drive signal is achieved, effectively solving the response deviation caused by the complex nonlinearity between the coil drive current, magnetic force, valve core displacement, and flow rate, as well as changes in operating conditions. At the same time, the online update of the dynamic threshold and the error closed-loop compensation mechanism significantly improve the robustness and long-term stability of the model to dynamic environmental changes, avoiding model aging and performance degradation. The closed-loop control strategy optimization based on deep reinforcement learning is further introduced, and the dynamic adjustment of the irrigation control amplitude and execution time is driven by real-time multi-source feedback data, realizing intelligent adaptation and refined irrigation management of the system, significantly improving the efficiency of water resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flow chart of an agricultural irrigation control method provided in an embodiment of the present application.

[0027] Figure 2 A schematic diagram of the structure of an agricultural irrigation control system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0029] Example 1, Figure 1 A flow chart of an agricultural irrigation control method provided in an embodiment of the present application includes the following steps:

[0030] S1, builds a crop water requirement model based on the fusion of crop physiological sensor data and data-driven models, and outputs the target irrigation flow in real time.

[0031] In this example, a data-driven model is constructed by collecting multi-source crop observation data and using intelligent algorithms such as machine learning, deep learning, and graph neural networks, without explicitly relying on crop water physiology equations. This data-driven model uses multi-source crop data as input, combined with a graph convolutional network and a Transformer time series network to form a hybrid graph-time series modeling network, serving as the backbone model for predicting crop water requirements.

[0032] This embodiment builds a crop water requirement model based on the fusion of crop physiological sensor data and a data-driven model, and outputs the target irrigation flow in real time. Specifically,

[0033] S11, collecting meteorological data, soil moisture data, and crop physiological index data of the target crop area to form a crop multi-source data set, wherein the meteorological data includes air temperature, air humidity, total solar radiation, wind speed and direction, and rainfall; the soil moisture data includes soil moisture content, soil temperature, soil electrical conductivity, soil type, and soil available water capacity; and the crop physiological index data includes leaf surface temperature, chlorophyll content, crop height, crop transpiration intensity, and leaf area index;

[0034] It should be noted that soil moisture data is the core data used to reflect the soil's water supply capacity, and crop physiological indicator data is used to reflect the water demand status of the crop itself and its dynamic response.

[0035] S12: Input the multi-source crop dataset into the time series prediction network that integrates graph convolution and Transformer to generate the basic prediction value of crop water requirement. The specific steps are as follows:

[0036] S121, define each target crop area as a topological map Nodes in , and extract the spatial relationship between adjacent plots based on the crop multi-source dataset to construct edge , and attach node attributes to each node , where node attributes include crop categories , soil characteristics , historical production and remote sensing characteristics ;

[0037] S122, assign weights to edges Constructing the graph structure , that is, the crop spatial relationship network, where is a node set, is the edge set, is the weighted edge matrix;

[0038] S123, extract spatial features of node attributes in the graph structure through graph convolutional neural network to obtain spatial feature embedding vector ;

[0039] S124: Obtain multi-source time series data for each target crop area, input the multi-source time series data into the Transformer time series modeling module, and extract time series features using the self-attention mechanism. ,Among them, the temporal features include time-dependent structural features,,stage features, attention focus weight features, multivariate coupling features,,and non-stationary variation features;

[0040] S125, fusing the spatial feature embedding vector with the time series feature to construct a crop water requirement prediction backbone model, and generating a basic crop water requirement prediction value based on the crop water requirement prediction backbone model.

[0041] The specific calculation formula of the weight is as follows:

[0042]

[0043] Where, For nodes and The weight between For nodes and The geographical distance between is the distance smoothing factor, is the node attribute similarity function, 、 For nodes and The node attributes between them.

[0044] It should be noted that the node attribute similarity function can use cosine similarity or RBF kernel function, and the distance smoothing factor can be understood as, for any two plots and , their spatial similarity is usually related to their geometric or ecological distance. If the distance is farther, the correlation should be moderately attenuated, so it is necessary to Mapping to graph edge weights , where the attenuation control term or mapping function is the distance smoothing factor, which is set based on experience.

[0045] The specific calculation formula of the spatial feature embedding vector is as follows:

[0046]

[0047] Where, is the spatial feature embedding vector, is the activation function, is the weighted edge matrix after adding self-connection, is the weighted edge matrix, for The corresponding degree matrix, It is a node attribute.

[0048] It should be noted that the activation function is the ReLU activation function, and the degree matrix is ​​a diagonal matrix, whose diagonal elements are the sum of the elements of each row of the adjacency matrix after adding self-connection.

[0049] The specific calculation formula of the main model for predicting crop water requirement is as follows:

[0050]

[0051] Where, is the basic predicted value of crop water requirement, is the spatial feature embedding vector, is the time series feature, is the fusion weight, is the bias term.

[0052] It should be noted that phase characteristics refer to the continuous trend of change within the crop growth stage, attention weight features refer to the weights of key influencing moments identified by the Transformer, multivariate coupling features refer to the cross-temporal interactions between variables such as temperature, humidity, and soil moisture, and non-stationary variation features refer to the dynamic changes in the mean, variance, and distribution of time series data over time. The backbone model for predicting crop water requirement forms a comprehensive forecast representation of crop water requirement by fusing the spatial feature embedding generated by graph convolution with the temporal features extracted by the Transformer. This can be understood as connecting a fully connected layer to the joint spatial-temporal embedding vector for regression output, resulting in the predicted water requirement value of the target plot at a unit time scale, which serves as the output of the prediction backbone model.

[0053] Furthermore, the introduction of a physiological sensor fusion mechanism into the core crop water requirement prediction model built using graph convolution and Transformer significantly enhances the model's real-time performance, personalization, robustness, interpretability, and decision-making support capabilities. Traditional data-driven models rely on historical statistical time series features and are inadequate for modeling changes in physiological responses caused by sudden meteorological variations, microscale water stress, transplanting, pests and diseases, and other factors. Physiological sensors (such as stem flow, leaf temperature, chlorophyll fluorescence, and stomatal conductance) can monitor crop water metabolism and stress responses in real time, capturing non-steady-state changes in crop water requirement caused by external disturbances. This enhances the model's ability to track and quickly correct sudden, short-term water fluctuations, improving system stability and robustness. Even within the same climate zone and soil type, crops can exhibit significant individual variations in water requirement curves and physiological responses due to differences in variety, management practices, planting density, and growth progression. The physiological sensor fusion mechanism enables a shift from regional average predictions to customized predictions at the individual / field level, providing physiological support for zoned, variable irrigation.

[0054] The time series prediction network that integrates graph convolution and Transformer can identify the coordinated changes in water demand in spatially similar areas, effectively supporting zoning control and differentiated irrigation regulation; Transformer has the ability to capture long-sequence global dependencies, can integrate the water response patterns of multiple historical seasons, and dynamically weight the input time series to improve the model's robustness to sudden drought and flood changes; graph convolution can process structured plots and crop information, and Transformer can process time series and high-dimensional meteorological variables. By jointly embedding node attributes and time series attributes, the model has good migration capabilities and is suitable for modeling water demand patterns in different climate zones, different soil types and crop varieties; GTCN-Transformer can be used as the backbone network and integrated with modules such as physiological sensor feedback modules and irrigation execution controllers to support end-to-end joint optimization, support subsequent closed-loop regulation with physiological stress feedback, and adaptive control strategies to build a complete "prediction-decision-execution" system.

[0055] S13, based on the real-time stem flow sensor and transpiration meter data, modifies the basic prediction value through an adaptive weighting mechanism and outputs the final value of crop water requirement; specifically:

[0056] S131, placing a stem flow sensor and a transpiration meter in the target area to obtain physiological indicator data, including but not limited to stem flow rate, stem flow rate change trend, transpiration rate, total transpiration, and stem water potential;

[0057] S132, obtaining an equivalent physiological feedback value based on normalization processing of the physiological indicator data, and performing differential calculation with a basic predicted value of crop water requirement to obtain a crop water stress residual data sequence;

[0058] S133, constructing a regression correction subnetwork based on the crop water stress residual data sequence, extracting a dynamic error trend, and obtaining a correction term output by the crop water requirement prediction backbone model through a gated weighted fusion mechanism;

[0059] S134, performing weighted fusion on the correction term and the output of the crop water requirement prediction main model to obtain a crop water requirement model.

[0060] Among them, the specific calculation formula of the equivalent physiological feedback value is as follows:

[0061]

[0062] Where, is the equivalent physiological feedback value, After normalization, Physiological index data, is the total number of physiological indicators, For the The weight coefficient of an indicator.

[0063] The specific calculation formula of the crop water stress residual data series is as follows:

[0064]

[0065] Where, is the crop water stress residual data series, is the equivalent physiological feedback value, is the basic predicted value of crop water requirement, is the feedback mapping coefficient.

[0066] It should be noted that the crop water stress residual data series refers to the time-varying difference between the predicted values ​​of the crop water requirement prediction backbone model within the current time window and the actual water stress levels reflected by the physiological sensors at the same time. The feedback mapping coefficient represents the mapping ratio of the equivalent physiological feedback value to the actual water requirement.

[0067] The specific calculation formula of the correction term is as follows:

[0068]

[0069] The specific calculation formula of the crop water requirement model is as follows:

[0070]

[0071] Where, is the target irrigation flow, is the basic predicted value of crop water requirement, is the gating coefficient, is the correction term, is the regression correction subnetwork, is the crop water stress residual data series.

[0072] It should be noted that the gating coefficient is calculated by the gating coefficient calculation expression in the gated weighted fusion mechanism, and is usually implemented by a sigmoid function.

[0073] A crop water requirement prediction backbone model is constructed through a time series prediction network that integrates graph convolution and Transformer. Feedback correction is performed based on a physiological sensor fusion mechanism, and the target irrigation flow rate is ultimately output. This model has the following advantages:

[0074] Structural spatiotemporal information fusion improves modeling accuracy: Graph convolutional networks can model spatial topological relationships within crop growing areas (such as field adjacency, irrigation pipe network transmission paths, and soil spatial heterogeneity), accurately expressing the spatial attributes of water demand characteristics; the Transformer model excels at processing long-distance temporal dependencies and can effectively explore the complex evolutionary patterns between crop historical meteorology, water dynamics, and crop status; the backbone model constructed by fusing the two has the dual modeling capabilities of spatial structure and temporal evolution, which is far superior to traditional methods based solely on statistics or simple time series models.

[0075] Introducing feedback on the actual physiological state of crops to enhance the model's true adaptability: Traditional irrigation models based on environmental factors or experience lack awareness of the current true water stress state of crops, which may lead to insufficient or wasted irrigation. By deploying physiological sensors such as leaf temperature, stem flow, and stomatal conductance, a feedback correction mechanism is constructed to obtain the true state expression of the crop's response to the external environment and irrigation. Dynamic compensation of the main model is performed through residual correction to ensure that the output irrigation amount is more in line with the actual water demand conditions, with greater robustness and field adaptability.

[0076] It has self-learning capabilities and can continuously optimize performance over time: the model architecture introduces a scalable feedback correction network and end-to-end training framework, enabling it to have continuous learning and adaptive updating capabilities; it can automatically adapt to unstable factors such as seasonal changes, climate fluctuations, sensor drift, and maintain long-term stable prediction performance.

[0077] S14, generating a target irrigation flow rate according to the final value.

[0078] S2, calculates the control error between the target irrigation flow and the actual collected flow, and generates the control signal of the proportional solenoid valve through the model predictive control strategy.

[0079] In this embodiment, in an intelligent irrigation control system, the step of generating proportional solenoid valve control signals using a model predictive control strategy based on the control error between the target irrigation flow rate and the actual collected flow rate is a key step in achieving precise water regulation, improving water use efficiency, and ensuring a stable crop water supply. The control error reflects the degree to which the current system deviates from the target. This provides the current error state to the MPC (Model Predictive Control) system. The control error drives the controller to generate correction signals; the larger the error, the more stringent the controller's adjustments, demonstrating adaptive characteristics.

[0080] The control error between the target irrigation flow rate and the actual collected flow rate is calculated, and the control signal of the proportional solenoid valve is generated through the model predictive control strategy, specifically:

[0081] S21, collecting the actual irrigation flow of the irrigation system in real time as the system output at the current moment, wherein the collecting device is a flow sensor installed in the pipeline, which is used to obtain the instantaneous flow value of the current irrigation unit;

[0082] S22, calculating the difference between the target irrigation flow rate and the actual irrigation flow rate to obtain a control error, wherein the control error is used to represent the degree of deviation between the current irrigation state and the target;

[0083] S23, obtaining historical irrigation system control input signals, flow rate responses, and external disturbance information, and constructing a dynamic prediction model for the irrigation system. The model is a differential equation model that can reflect the temporal relationship between the proportional solenoid valve control signal and the irrigation flow rate, and is used to model the flow rate change trend within multiple future prediction steps;

[0084] S24, constructing an objective function for minimizing the irrigation control error and the control signal variation amplitude within a future prediction time domain based on the dynamic prediction model and the control error, wherein the objective function includes a control error and a control smoothness term, and defines constraints;

[0085] S25, using the rolling time domain optimization method, solves the optimal solution of the objective function in each control cycle, obtains the optimal control input in the future prediction time domain, and selects the control quantity at the current moment as the control signal of the proportional solenoid valve.

[0086] Among them, the objective function, the specific calculation formula is as follows:

[0087]

[0088] Where, is the objective function, is the prediction step index, where =1,2,3,..., , To predict the time domain length, For the The target irrigation flow at the moment, To collect actual traffic, For the control input signal, is the control input at the previous moment, is the control smoothing weight factor.

[0089] Among them, the constraints include control input boundary constraints and control input change rate constraints;

[0090] Among them, the control input boundary constraints are:

[0091]

[0092] Control input change rate constraint:

[0093]

[0094] Where, and The minimum and maximum control signal values ​​allowed by the proportional solenoid valve to prevent overflow control. The maximum adjustment variation allowed by the system limits the valve execution frequency and mechanical wear to ensure smooth and reliable system operation.

[0095] It should be noted that the control input signal is the actuator signal for controlling the proportional solenoid valve. A larger control smoothing weight factor results in a smoother control signal, resulting in a slower but more robust response. A smaller control smoothing weight factor results in a faster tracking of the target flow rate, resulting in a more responsive but potentially unstable system. To control the error, is the control smoothness term. The core principle of the MPC control strategy is to execute only the first value in the optimal control sequence as the control signal at the current moment, then discard the remaining value and re-predict it in the next control cycle.

[0096] S3, based on the initial static mapping model fitted by the Sigmoid function and the dynamic correction model driven by state perception, converts the control signal of the proportional solenoid valve into the driving signal of the proportional solenoid valve.

[0097] In this embodiment, the control signals are mostly normalized signals or desired quantities, such as target flow, opening, and position. These are abstract desired values. However, proportional solenoid valves, which rely on current or voltage excitation to generate magnetic force to propel the valve core, must be converted into real physical drive quantities. Furthermore, the control signal should have a continuously adjustable, approximately linear relationship with the valve core position, flow, or pressure. However, a complex, nonlinear chain exists between coil drive current, magnetic force, displacement, and flow. Therefore, a nonlinear mapping model from control signal to drive current is necessary to accurately translate control intent into coil action and achieve proportional regulation.

[0098] The initial static mapping model based on Sigmoid function fitting and the dynamic correction model driven by state perception convert the control signal of the proportional solenoid valve into the drive signal of the proportional solenoid valve, specifically:

[0099] S31, acquiring historical driving data of the proportional solenoid valve under different control signals, constructing an initial nonlinear mapping relationship between the control signal and the driving signal, and fitting the relationship based on a Sigmoid saturation function to obtain an initial static mapping model, wherein the historical driving data includes historical voltage and current driving data;

[0100] S32, acquiring key state variables under the solenoid valve working environment and constructing a state perception variable set, wherein the key state variables include but are not limited to coil temperature, oil temperature or viscosity, load pressure, and motion wear state;

[0101] S33, constructing a dynamic correction model through a recurrent neural network based on the state perception variable and first data, where the first data includes a control signal and a drive response;

[0102] S34, using the initial static mapping model as a feedforward module to perform preliminary nonlinear mapping on the control signal, output a fitting drive signal, and output a dynamic response offset based on the dynamic correction model;

[0103] S35, by superimposing the dynamic response offset and the fitted driving signal, the final driving response output is obtained, and the model is updated and dynamically corrected online based on the dynamic threshold to achieve dynamic environment adaptation and error closed-loop compensation;

[0104] S36, the constructed model is integrated into the proportional solenoid valve control system to achieve high-precision and dynamic response consistent conversion of control signal to drive signal.

[0105] The specific calculation formula of the initial static mapping model is as follows:

[0106]

[0107] Where, is the fitting driving signal under static mapping, is the saturation amplitude parameter, corresponding to the maximum drive response, Input control signal for proportional solenoid valve, is the Sigmoid center point, is the slope of the Sigmoid function.

[0108] The specific calculation formula of the dynamic correction model is as follows:

[0109]

[0110] Where, is the dynamic response offset, is a state-aware variable, Input control signal for proportional solenoid valve, is the neural network mapping function.

[0111] The final drive response output is calculated as follows:

[0112]

[0113] Where, is the final drive response output, is the dynamic response offset, Fitting driving signal under static mapping.

[0114] It should be noted that the drive response refers to the electrical drive signal (e.g., current / voltage) output by the controller and applied to the proportional solenoid valve coil at the kth control cycle. The drive response of a proportional solenoid valve actually consists of two components: a static component and a dynamic disturbance. The static component corresponds to the nonlinear input-output characteristic under ideal operating conditions (e.g., calibration temperature, standard oil, and no aging). The dynamic disturbance reflects the actual response offset caused by changes in state variables (e.g., temperature rise, wear, load changes, etc.). The final drive response output is obtained by superimposing the dynamic response offset with the fitted drive signal. This can be understood as separating the static nonlinear characteristics and dynamic disturbance response in the control mapping, thereby achieving a drive signal generation process with greater structural clarity, computational efficiency, and model adaptability. The dynamic correction model uses a recurrent neural network mapping function to model the combined influence of the control signal and state-sensing variables and outputs a dynamic offset.

[0115] Online updates and dynamic corrections can improve the system's ability to adapt to environmental changes. The online update mechanism can identify sudden changes in model errors in real time, make timely corrections to the model, and enhance the robustness and adaptability of the control system to non-stationary environments. It can effectively suppress the risks of model aging and drift. Through sliding window error analysis + incremental fine-tuning, model aging failure can be avoided and the model life cycle can be extended. The risk of false operation and reduced drive efficiency due to model mismatch is reduced. Dynamic correction can significantly reduce the drive signal prediction error and improve the consistency between actuator behavior and target control signal, thereby ensuring system safety and execution accuracy.

[0116] The online updating and dynamic correction of the model based on the dynamic threshold to achieve dynamic environment adaptation and error closed-loop compensation are specifically as follows:

[0117] S351, collecting second data and calculating the current driving response error, wherein the second data includes a real-time control signal, a model output, and an actual driving response;

[0118] S352: Build an error sliding window and make a judgment based on the error mean and change trend, triggering an online update based on the judgment result, wherein the judgment includes whether the error exceeds a preset threshold or shows a mutation trend;

[0119] S353, after triggering the online update, introducing a structure supporting online fine-tuning into the initial static mapping model and the dynamic correction model, wherein the initial static mapping model adopts a multi-layer perceptron structure capable of incremental updates, and the dynamic correction model adopts a recurrent neural network supporting local parameter fine-tuning;

[0120] S354, extracting model training data within the current time window to construct a training set and a validation set, wherein the model training data includes control signals, state perception variable sequences, and actual drive response data;

[0121] S355, performing a finite-step gradient update on the model parameters based on the training set using a preset micro-batch training strategy;

[0122] S356: After the update is completed, the model errors before and after the update are compared based on the validation data set. If the error reduction ratio after the update exceeds the preset ratio threshold, the new model is retained; otherwise, the model parameters are rolled back to the previous optimal model.

[0123] It should be noted that the mutation trend includes the error continuously increasing in the sliding time window, the error growth rate exceeding a threshold, and the error variance or the maximum and minimum value difference significantly increasing (such as exceeding two standard deviations of the mean).

[0124] S4 collects multi-dimensional feedback data of valve execution and performs closed-loop comparison between the driving signal of the proportional solenoid valve and the feedback data, and adaptively optimizes the irrigation control strategy through the reinforcement learning controller.

[0125] In this embodiment, by continuously collecting multi-dimensional feedback data, dynamically executing strategy iterations and modifying irrigation control strategies, a fully closed-loop control evolution process of control signal generation - execution - environmental feedback - strategy optimization is formed, thereby achieving the goal of enhancing the adaptive capability of the irrigation control system and optimizing long-term water conservation.

[0126] Collect multi-dimensional feedback data from valve execution and perform a closed-loop comparison between the proportional solenoid valve's drive signal and the feedback data. Adaptively optimize the irrigation control strategy through a reinforcement learning controller. Specifically:

[0127] S41, during the irrigation process after the proportional solenoid valve executes the drive signal, obtaining multi-dimensional feedback data directly related to its response effect, the multi-dimensional feedback data including physiological response data, soil moisture change data in the soil surface layer and root zone of the irrigation area, irrigation energy consumption penalty, and water use efficiency. The physiological response data is calculated based on crop leaf temperature, stem electrical conductivity, chlorophyll content, and photosynthetic rate. The soil moisture change data is calculated based on soil moisture content and target moisture content.

[0128] S42, building a reinforcement learning controller based on a deep reinforcement learning policy optimization network;

[0129] S43, dynamically optimizing irrigation control strategy parameters according to the output result of the reinforcement learning controller, wherein the control strategy parameters include the control amplitude of the proportional solenoid valve and the actual irrigation execution duration.

[0130] In step S43, dynamic optimization is performed, specifically:

[0131] Dynamically adjust the control amplitude of the proportional solenoid valve to achieve fine-grained flow control;

[0132] Adjust the actual irrigation execution duration to achieve time distribution control of total water volume;

[0133] Reinforcement learning controller, specifically:

[0134] A state space composed of a multidimensional state vector, wherein the multidimensional state vector includes the control signal at the previous moment, the irrigation duration, the current drive signal, and environmental feedback variables, wherein the environmental feedback variables include the current soil moisture content, the current crop physiological state indicator set, the current wind speed, and the ambient temperature;

[0135] An action space composed of control strategy parameters;

[0136] Constructing a reward function for the target irrigation effect through multi-dimensional feedback data, the reward function includes physiological response data, soil moisture change data, irrigation water use efficiency and irrigation energy consumption penalty;

[0137] Construct a reinforcement learning controller based on the state space, action space, and reward function;

[0138] The reinforcement learning controller outputs the action with the maximum reward function in the current state to obtain the optimal control strategy parameters;

[0139] The DDPG optimization strategy method is used to update the reinforcement learning controller, specifically:

[0140] Calculate the immediate reward function based on the multi-dimensional feedback data after the action is performed;

[0141] Evaluate state-action value through the Critic network;

[0142] Perform policy gradient ascent optimization through the Actor network;

[0143] Soft updates to reinforcement learning controllers for policy iteration.

[0144] In step S41, the physiological response data is calculated using the following formula:

[0145]

[0146] Where, is the physiological response data, is the total number of physiological parameters, For the Physiological parameters (such as crop leaf temperature, stem conductivity, chlorophyll content and photosynthetic rate) The observed value of is the observation value of the previous time step, 、 are the maximum and minimum reference values ​​used for normalization.

[0147] The specific calculation formula for the soil moisture change data is as follows:

[0148]

[0149] Where, is the soil moisture change data, is the current soil moisture content, Target moisture content set according to crop growth stage

[0150] The specific calculation formula of the irrigation energy consumption penalty is as follows:

[0151]

[0152] Where, Penalty for irrigation energy consumption, is the energy consumption penalty coefficient, is the opening control signal value of the proportional solenoid valve at the current moment, The current irrigation duration.

[0153] The specific calculation formula of the reward function is as follows:

[0154]

[0155] Where, is the reward function, is the physiological response data, For irrigation water use efficiency, is the soil moisture change data, Penalty for irrigation energy consumption, 、 、 、 are weight coefficients respectively.

[0156] It should be noted that in each decision cycle, the controller collects three types of information, namely crop, soil and environment, and historical control behaviors, constructs the current state vector, and uses deep policy optimization methods (such as DDPG and PPO) to construct a policy function, inputs the state vector, and outputs the action vector; according to the output action of the policy network, the system performs the following control operations: updates the solenoid valve control signal, sets the duration of the current irrigation cycle, and implements physical layer irrigation behavior; after executing the irrigation behavior, the system evaluates the comprehensive benefits of this round of decision-making in real time through the reward function, and feeds back to the policy network to guide learning and complete iterative optimization.

[0157] It should be noted that the closed-loop comparison of the proportional solenoid valve's drive signal and feedback data can be understood as calculating the reward value based on the feedback data. The system continuously compares and evaluates the irrigation energy consumption penalty after executing the drive signal with the target effect, and uses this evaluation result (reward) to drive the optimization of the control strategy and generate a new drive signal, forming a complete, adaptive closed-loop control system.

[0158] Example 2, Figure 2 This is a schematic diagram of the structure of an agricultural irrigation control system provided in an embodiment of the present application, including a target irrigation flow acquisition module, a control signal generation module, a control signal conversion module, and a control strategy feedback optimization module. There are connections between the modules:

[0159] The target irrigation flow acquisition module is used to build a crop water requirement model based on the integration of crop physiological sensor data and data-driven models, and output the target irrigation flow in real time;

[0160] The control signal generation module is used to calculate the control error between the target irrigation flow and the actual collected flow, and generate the original control signal of the proportional solenoid valve through the model predictive control strategy;

[0161] The control signal conversion module is used to input the original control signal into the static mapping model fitted by the Sigmoid function, superimpose the dynamic correction model output driven by the state perception data, and generate the final drive signal of the proportional solenoid valve;

[0162] The control strategy feedback optimization module is used to collect valve execution feedback data, perform closed-loop comparison between the drive signal and the feedback data, and adaptively optimize the irrigation control strategy parameters through a reinforcement learning controller.

[0163] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0164] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0165] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0166] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0167] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0168] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An agricultural irrigation control method, characterized in that: The steps include: A crop water requirement model is constructed based on the integration of crop physiological sensor data and data-driven models, and the target irrigation flow is output in real time; The data-driven model uses multi-source crop data as input, and then fuses the graph convolutional network and the Transformer time series network to form a graph-time series hybrid modeling network as the backbone model for crop water requirement prediction. The graph convolutional network and the Transformer temporal network are combined to form a graph-temporal hybrid modeling network, specifically: The target crop area is defined as a node in the topological graph, and a crop spatial relationship network is constructed. The spatial feature embedding vector of the node is extracted through a graph convolutional neural network. Obtain multi-source time series data for each crop region, input it into the Transformer time series modeling module, and use the self-attention mechanism to extract time series features; Fusion of spatial feature embedding vectors and temporal features to generate basic predictions of crop water requirements; Calculate the control error between the target irrigation flow and the actual collected flow, and generate the original control signal of the proportional solenoid valve through the model predictive control strategy; The original control signal is input into the static mapping model fitted by the Sigmoid function, and the dynamic correction model output driven by the state perception data is superimposed to generate the final drive signal of the proportional solenoid valve; The valve execution feedback data is collected, the drive signal is compared with the feedback data in a closed loop, and the irrigation control strategy parameters are adaptively optimized through the reinforcement learning controller.

2. The agricultural irrigation control method according to claim 1, characterized in that: The crop water requirement model is constructed based on the fusion of crop physiological sensor data and data-driven model, and the target irrigation flow is output in real time. Specifically: Collect meteorological data, soil moisture data, and crop physiological index data of the target crop area to form a multi-source crop data set; Input multi-source crop data into a time series prediction network that integrates graph convolution and Transformer to generate a basic prediction value of crop water requirement; Based on real-time stem flow sensor and transpiration meter data, the basic prediction value is corrected through an adaptive weighting mechanism to output the final value of the crop water requirement model; Generate target irrigation flow based on the final value.

3. The agricultural irrigation control method according to claim 2, characterized in that: Based on the real-time stem flow sensor and transpiration meter data, the basic prediction value is corrected through an adaptive weighting mechanism to output the final value of the crop water requirement model, specifically: Acquire physiological index data and generate equivalent physiological feedback values ​​after normalization; Calculating the difference between the equivalent physiological feedback value and the basic predicted value of crop water requirement to generate a crop water stress residual data series; Constructing a regression correction sub-network based on the residual data sequence, and outputting a correction term through a gated weighted fusion mechanism; The correction term is weighted and fused with the basic predicted value of crop water requirement to generate the output value of the crop water requirement model.

4. The agricultural irrigation control method according to claim 3, characterized in that: The control error between the target irrigation flow and the actual collected flow is calculated, and the original control signal of the proportional solenoid valve is generated through the model predictive control strategy, specifically: Collect the actual irrigation flow of the irrigation system in real time as the current system output; Calculate the control error between the target irrigation flow and the actual irrigation flow; Obtain historical irrigation system control input signals, flow response, and external disturbance information to establish a dynamic prediction model for the irrigation system; Based on the dynamic prediction model and control error, construct the objective function and set the system constraints; The objective function is solved by using the rolling horizon optimization method to generate the optimal control sequence in the future horizon; The first instruction of the optimal control sequence is extracted as the control signal of the proportional solenoid valve.

5. The agricultural irrigation control method according to claim 4, characterized in that: The original control signal is input into the static mapping model fitted by the Sigmoid function, and the dynamic correction model output driven by the state perception data is superimposed to generate the final drive signal of the proportional solenoid valve, specifically: Obtain historical driving data, build a nonlinear mapping relationship between the original control signal and the driving signal, and generate an initial static mapping model based on Sigmoid saturation function fitting; Real-time collection of key state variables in the solenoid valve working environment and construction of a state-aware variable set; Constructing a dynamic correction model through a recurrent neural network based on a set of state-sensing variables and first data, wherein the first data includes a control signal and a drive response; The initial static mapping model is used as a feedforward module to output the fitted driving signal, while the dynamic response offset is output through the dynamic correction model. The dynamic response offset is superimposed on the fitted driving signal to generate the final driving response output; The dynamic model is updated online and corrected in real time through a preset dynamic threshold.

6. The agricultural irrigation control method according to claim 5, characterized in that: The online updating and real-time correction of the dynamic model by presetting the dynamic threshold is specifically as follows: Collecting second data and calculating a current driving response error, wherein the second data includes a real-time control signal, a model output, and an actual driving response; Build an error sliding window and determine whether to trigger an online update based on the error mean and change trend; After triggering the online update, a structure supporting online fine-tuning is introduced into the initial static mapping model and the dynamic correction model. The structure includes a multi-layer perceptron structure capable of incremental updates for the initial static mapping model and a recurrent neural network supporting local parameter fine-tuning for the dynamic correction model. Extracting model training data within the current time window to construct a training set and a validation set, wherein the model training data includes control signals, state perception variable sequences, and actual drive response data; Based on the training set, a finite-step gradient update is performed on the model parameters using a preset micro-batch training strategy. After the update is completed, the model errors before and after the update are compared based on the validation data set. If the error reduction ratio after the update exceeds the preset ratio threshold, the new model is retained; otherwise, it rolls back to the previous optimal model parameters.

7. The agricultural irrigation control method according to claim 6, characterized in that: The valve execution feedback data is collected, the drive signal is compared with the feedback data in a closed loop, and the irrigation control strategy parameters are adaptively optimized through the reinforcement learning controller, specifically: During the irrigation process after the proportional solenoid valve executes the driving signal, multi-dimensional feedback data directly related to the irrigation effect is collected in real time; Build a deep reinforcement learning policy optimization network to generate a reinforcement learning controller; According to the output of the reinforcement learning controller, the control strategy parameters of the proportional solenoid valve are dynamically optimized, and the control strategy parameters include the control amplitude of the proportional solenoid valve and the actual irrigation execution duration.

8. The agricultural irrigation control method according to claim 7, characterized in that: The reinforcement learning controller is specifically: Construct a state space consisting of multi-dimensional state vectors and an action space consisting of control strategy parameters; Constructing a reward function for the target irrigation effect through multi-dimensional feedback data, the reward function includes physiological response data, soil moisture change data, irrigation water use efficiency and irrigation energy consumption penalty; Construct a reinforcement learning controller based on the state space, action space, and reward function; The reinforcement learning controller outputs the action with the maximum reward function in the current state to obtain the optimal control strategy parameters.

9. A system using the agricultural irrigation control method according to any one of claims 1 to 8, characterized in that: It includes target irrigation flow acquisition module, control signal generation module, control signal conversion module and control strategy feedback optimization module. There are connections between modules: The target irrigation flow acquisition module is used to build a crop water requirement model based on the integration of crop physiological sensor data and data-driven models, and output the target irrigation flow in real time; The control signal generation module is used to calculate the control error between the target irrigation flow and the actual collected flow, and generate the original control signal of the proportional solenoid valve through the model predictive control strategy; The control signal conversion module is used to input the original control signal into the static mapping model fitted by the Sigmoid function, superimpose the dynamic correction model output driven by the state perception data, and generate the final drive signal of the proportional solenoid valve; The control strategy feedback optimization module is used to collect valve execution feedback data, perform closed-loop comparison between the drive signal and the feedback data, and adaptively optimize the irrigation control strategy parameters through a reinforcement learning controller.

Images