A brackish water desalination closed-loop AI precise water and fertilizer circulation hydroponics treatment system

Abstract

This invention discloses a closed-loop AI-based precision water and fertilizer recycling hydroponic treatment system for brackish water desalination, applied in the field of intelligent control and digital twin control technology for facility agriculture. It includes a sensing module, a computing module, a decision-making module, and a self-learning module. The sensing module acquires multi-source state feature data, cleans and aligns it to output a state vector; the computing module constructs resource, environmental, and health feature matrices, and after dimensionality reduction, calculates the dynamic weight allocation for each business dimension based on Shannon information entropy; the decision-making module combines dynamic weights, state vectors, and weather forecasts to construct multi-objective evaluation constraint equations, and optimizes and issues the best control commands; the self-learning module collects actual response data, calculates the mean square deviation, and fine-tunes the weight gain and bias compensation factor of the decision objective function online when exceeding the time limit. This invention eliminates the reliance on complex prediction models and achieves high-precision, low-computing-power, adaptive closed-loop precision water and fertilizer control in brackish water.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and multi-physics field collaborative optimization technology in facility agriculture, specifically to a brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system. Background Technology

[0002] In the high-salinity and arid regions of Northwest my country, using brackish water for hydroponics is a crucial way to address agricultural water scarcity. However, the process of brackish water desalination and greenhouse water-fertilizer circulation involves the coordinated regulation of multiple variables, including desalination pump pressure, water-fertilizer return ratio, nutrient solution concentration (EC / pH), and microclimate. Traditional hydroponic systems often rely on fixed thresholds or manual experience for control. When faced with complex conditions such as drastic fluctuations in brackish water salinity and sudden weather changes, these systems suffer from problems such as delayed water and fertilizer regulation, high desalination energy consumption, low resource recovery rates, and extreme susceptibility of crops to salt damage.

[0003] To improve control precision, digital twin and AI predictive control technologies have been gradually introduced into facility agriculture in recent years. However, existing AI solutions heavily rely on complex high-order physical equations (such as fluid dynamics CFD and Richards' water transport model) or deep learning algorithms (such as LSTM), requiring expensive high-performance edge servers and incurring extremely high training and maintenance costs, making them very difficult to promote in agricultural fields. Furthermore, existing control systems lack the ability to dynamically correct for physical hardware degradation such as reverse osmosis membrane fouling and blockage, energy storage battery aging, and sensor zero-point drift, leading to severe steady-state control deviations after long-term operation. Therefore, developing a brackish water and fertilizer precision circulation control system with low computational overhead, high robustness, and adaptive closed-loop correction capabilities is of great significance.

[0004] Therefore, the core technical challenge to achieve high-precision, adaptive, high-yield, and stable production in brackish water facility agriculture lies in how to establish a unified abstract framework for agricultural operations, transform complex underlying algorithm indicators into resource, environmental, and health control semantics that are crucial to crop production, and on this basis, build a closed-loop digital twin control system with forward-looking predictive capabilities. Summary of the Invention

[0005] The purpose of this application is to address at least one of the aforementioned technical defects, particularly the technical defects in the prior art, such as the inability of discrete control to coordinate, the disconnect between optimization algorithm indicators and agricultural business semantics, the lack of forward-looking rolling prediction, and the inability of the system to adaptively evolve, which lead to the waste of water, fertilizer, and energy resources and severe wear and tear on core equipment.

[0006] In the first aspect, this application provides a brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system, including a sensing module, a computing module, a decision-making module and a self-learning module; The sensing module is configured with a multi-dimensional sensor network to acquire multi-source physical state feature data during the operation of the greenhouse, and to filter, clean and align the data in time and space to output a state data vector. The calculation module is used to construct a resource matrix, an environment matrix, and a health matrix based on the state data vector. After normalization, it is used to reduce the dimensionality based on the principal component analysis algorithm and calculate the discrete probability of the state distribution. The normalized information entropy is calculated by substituting it into the Shannon information entropy formula. Based on the normalized information entropy, the dynamic allocation weights of resources, environment, and health are obtained. The decision-making module is used to construct a multi-objective evaluation constraint equation, including resource utilization, environmental adaptability, plant health, system stability and operational safety, based on the dynamic allocation weights of resources, environment, and health, as well as the current state data vector and meteorological forecast data. It then optimizes and outputs the optimal control command and issues it for execution. The self-learning module is used to collect actual environmental response data after the execution of instructions, calculate the deviation statistics between the actual environmental response data and the expected target setting value, and send a parameter correction signal to the decision module when the deviation statistics are greater than the preset limit, so as to fine-tune the target weights and control gain parameters in the multi-objective evaluation constraint equation online.

[0007] As an optional implementation, the multi-dimensional sensor network includes a brackish water ion field sensor matrix, an environmental field climate sensor matrix, an energy operation status sensor matrix, and a crop canopy image multispectral acquisition device; The multi-source physical state characteristic data includes brackish water ion field data, environmental field climate data, energy operation data, and crop physiological phenotypic data; The multi-source physical state feature data is packaged and encapsulated.

[0008] As an optional implementation, the filtering, cleaning, and spatiotemporal alignment processing logic specifically includes: By using the sliding unscented Kalman filter algorithm, the continuous analog electrical signals of environmental field climate data and energy operation data in each dimension are estimated, and noise and electromagnetic pulse noise in the sensor acquisition process are filtered out. By using high-order Lagrange polynomial interpolation, brackish water ion field data with different acquisition frequencies are corrected and aligned to a uniform 1-hour clock step, and the equivalent value of brackish water ion field data at the standard control time is calculated. The irregularly distributed crop physiological phenotypic data are interpolated and reconstructed onto a unified greenhouse 3D grid node using the Kriging spatial interpolation algorithm. All multi-source physical state feature data that have been filtered, cleaned and spatiotemporally aligned are uniformly normalized and then one-dimensionally concatenated to obtain a state data vector.

[0009] As an optional implementation, the calculation module is used to extract state data vectors from the current moment to the past 24 hours to construct a resource matrix, an environment matrix, and a health matrix; The row vectors of the resource matrix are, in order, the desalination system water production efficiency, reverse osmosis membrane recovery rate, photovoltaic utilization rate, water and fertilizer recycling recovery rate, and root zone conductivity variation variance. The row vectors of the environmental matrix are, in order: air temperature fluctuation rate, air relative humidity fluctuation rate, photosynthetically active radiation intensity, saturated vapor pressure difference, weather forecast temperature deviation, and root zone dissolved oxygen fluctuation rate. The row vectors of the health matrix are, in order, the relative distribution of chlorophyll, the leaf area index of the whole plant, the root length, the fruit sugar content, and the plant disease index.

[0010] As an optional implementation, the data processing logic of the computing module includes: The maximum-minimum-extreme-boundary normalization is performed on each row vector of the resource matrix, environment matrix, and health matrix to obtain dimensionless resource feature matrix, environment feature matrix, and health feature matrix. Principal component analysis is used to project each feature matrix onto the corresponding two-dimensional principal feature projection space, and two-dimensional discrete grid cells are divided in the two-dimensional principal feature projection space. The probability of sample points falling into each grid cell is statistically analyzed to obtain the discrete probability of the state distribution corresponding to each feature matrix. Based on Shannon's information entropy formula, the basic information entropy of the resource matrix, environment matrix, and health matrix are calculated respectively, and normalized by combining the historical maximum information entropy recorded in historical data to obtain normalized resource information entropy, normalized environment information entropy, and normalized health information entropy. Calculate the proportion of each normalized information entropy in the sum of the three to obtain the dynamic allocation weights of resources, environment, and health.

[0011] As an optional implementation, the multi-objective evaluation constraint equation includes a sub-objective function for maximizing resource utilization, a sub-objective function for optimal environmental adaptation, a sub-objective function for maximizing plant health, a sub-objective function for maximizing system stability, and a sub-objective function for maximizing operational safety, the mathematical expressions of which are: ; , ; in, For multi-objective evaluation constraint equations, Candidate control vectors, For the sub-objective function of maximizing resource utilization, The optimal environmental adaptation sub-objective function. For the maximum plant health sub-objective function, The objective function is the sub-objective function for maximizing system stability. For the maximum operational safety sub-objective function, Dynamically assign weights to resources. Dynamically assign weights to the environment. Dynamically assign weights to health status. As the weight of the equipment's stable state, As meteorological disaster risk weights, That is , That is , They are respectively The final dynamic weights after normalization.

[0012] As an optional implementation, the maximum resource utilization sub-objective function is used to calculate the product of the proportion of total water and fertilizer reuse in the total input raw materials and the photovoltaic power generation self-sufficiency rate; The optimal environmental adaptation sub-objective function is used to calculate the deviation of the current air temperature and relative humidity from the suitable growth target value; The maximum plant health sub-objective function is used to calculate the deviation of the current leaf area index and the relative chlorophyll distribution from the optimal reference value.

[0013] As an optional implementation, the maximum system stability sub-objective function is used to apply action penalties to the single-step adjustment of the reverse osmosis system operating pressure and the single-step change in the energy storage battery discharge depth. The maximum operational safety sub-objective function is used to calculate the operational safety protection level under external severe weather conditions based on the rainfall probability and solar radiation intensity predicted by the weather forecast.

[0014] As an optional implementation, the decision module iteratively optimizes the multi-objective evaluation constraint equations in the constraint space using an adaptive multi-objective particle swarm optimization algorithm at each control clock step, solves for the optimal control vector, and issues it for execution. The control vectors include: desalination pump pressure, reflux ratio, irrigation EC and pH, LED supplemental lighting power, fan speed and heating power.

[0015] As an optional implementation, after the control action is executed, the self-learning module collects the first and second actual physical measurement values ​​of the sensor at the next moment, and calculates the mean square deviation (MSE) between the actual physical measurement values ​​and the target set value. When the calculated mean square deviation (MSE) exceeds the preset limit, a dynamic fine-tuning procedure is triggered to adjust the bias gain of the corresponding sub-target in the decision module online, thereby achieving closed-loop adaptive correction.

[0016] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: This system constructs a resource matrix, an environment matrix, and a health matrix, combines principal component analysis algorithm and Shannon information entropy to calculate the normalized information entropy of each business dimension, and dynamically generates dynamic resource allocation weights, dynamic environment allocation weights, and dynamic health allocation weights based on the normalized information entropy. This enables the target weights in the multi-objective evaluation constraint equations to be adjusted in real time according to the system's operating status, realizing the correspondence between mathematical evaluation indicators and changes in agricultural business status, and improving the interpretability and adaptability of the multi-objective decision-making process.

[0017] This system directly constructs multi-objective evaluation constraint equations based on state data vectors and meteorological forecast data. It comprehensively considers resource utilization, environmental adaptability, plant health, system stability, and operational safety. The optimal control vector is obtained through an adaptive multi-objective particle swarm optimization algorithm, which avoids the computational overhead caused by complex models. While ensuring control accuracy, it reduces the system's computational complexity and facilitates engineering deployment.

[0018] This system uses principal component analysis to extract the main changing features of resource, environmental, and health feature matrices. It objectively evaluates the operational status of each business dimension through discrete grid probability statistics and Shannon information entropy calculation. When the fluctuation of a certain business dimension increases, its corresponding normalized information entropy and dynamically allocated weights increase simultaneously, enabling the decision-making module to automatically increase the influence of the corresponding objective function in the overall evaluation and reduce the inadequacy of fixed weight configuration for different operating conditions.

[0019] This system incorporates a self-learning module. After the control action is executed, based on the mean square deviation between the actual physical measurement value and the target setpoint, it triggers both fast-change and slow-change dynamic fine-tuning programs. Employing a proportional adaptive update method, it online corrects the temperature compensation correction coefficient, humidity compensation correction coefficient, resource compensation correction coefficient, leaf area index reference target value, chlorophyll relative distribution content reference target value, and health dynamic allocation weights. The updated parameters are then used in the calculation of the multi-objective evaluation constraint equation for the next control cycle, achieving closed-loop online adaptive fine-tuning. This dynamically compensates for control deviations caused by factors such as equipment aging, changes in water source quality, and membrane scaling, improving the system's stability and robustness. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 The flowchart illustrates a closed-loop AI-based precision water and fertilizer recycling hydroponic treatment system for brackish water desalination, as provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Furthermore, the technical features in the various embodiments or individual embodiments provided in this application can be arbitrarily combined with each other to form a feasible technical solution. Such combination is not constrained by the order of steps and / or the structural composition mode, but must be based on the ability of those skilled in the art to implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0024] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described below are merely illustrative. For example, the division of units and modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or modules can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, and can be electrical, mechanical or other forms.

[0025] In addition, each functional unit in the various embodiments of this application can be integrated into a single processor, or each unit can be a separate device, or two or more units can be integrated into a single device; each functional unit in the various embodiments of this application can be implemented in hardware or in the form of hardware plus software functional units.

[0026] Those skilled in the art will understand that all or part of the steps of the following method embodiments can be implemented by program instructions and related hardware. The aforementioned program instructions can be stored in a computer-readable storage medium. When executed, the program instructions perform the steps of the following method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks. It should be understood that the use of terms such as "system," "device," "unit," and / or "module" in this application is merely to distinguish different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they can be replaced by other expressions.

[0027] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature marked "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" or "several" means two or more, unless otherwise explicitly specified. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes and to aid those skilled in the art, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and objectives of this application, should still fall within the scope of the technical content disclosed in this application.

[0028] If a flowchart is used in this application, it is used to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0029] It should also be noted that, in this document, terms such as “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the article or device that includes the aforementioned element.

[0030] like Figure 1As shown, a brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system includes a sensing module, a computing module, a decision-making module and a self-learning module. The sensing module is configured with a multi-dimensional sensor network to acquire multi-source physical state feature data during the operation of the greenhouse, and to filter, clean and align the data in time and space to output a state data vector. The calculation module is used to construct a resource matrix, an environment matrix, and a health matrix based on the state data vector. After normalization, it is used to reduce the dimensionality based on the principal component analysis algorithm and calculate the discrete probability of the state distribution. The normalized information entropy is calculated by substituting it into the Shannon information entropy formula, and the dynamic allocation weights of resources, environment, and health are obtained based on the normalized information entropy. The decision-making module is used to construct a multi-objective evaluation constraint equation, including resource utilization, environmental adaptability, plant health, system stability and operational safety, based on the dynamic allocation weights of resources, environment, and health, as well as the current state data vector and weather forecast data. It then optimizes and outputs the optimal control command and issues it for execution. The self-learning module is used to collect actual environmental response data after executing instructions, calculate the deviation statistics between the actual environmental response data and the expected target set value, and send parameter correction signals to the decision module when the deviation statistics exceed the preset limit, so as to fine-tune the target weights and control gain parameters in the multi-objective evaluation constraint equation online.

[0031] This invention transforms multidimensional physical states into dynamic business allocation weights through an information entropy algorithm, enabling adaptive tilting of control computing power and execution resources towards high-risk business dimensions. It abandons the reliance on complex prediction models and adopts lightweight multi-objective decision-making based on real-time status and weather forecasts, greatly reducing the computational complexity of the algorithm and the cost of hardware deployment. Combined with a closed-loop self-learning fine-tuning mechanism based on mean square deviation, it can dynamically compensate for steady-state deviations caused by equipment aging, membrane fouling, and sudden changes in the external environment online, achieving high-precision, robust, and adaptive closed-loop precise control of brackish water desalination and hydroponic facilities with low computing power overhead.

[0032] The multi-dimensional sensor network includes a brackish water ion field sensor matrix, an environmental field climate sensor matrix, an energy operation status sensor matrix, and a crop canopy image multispectral acquisition device; Multi-source physical state characteristic data include brackish water ion field data, environmental field climate data, energy operation data, and crop physiological phenotypic data; Multi-source physical state feature data are packaged and encapsulated.

[0033] Specifically, the brackish water ion field sensor matrix is ​​used to collect brackish water ion field data, including nutrient solution reflux conductivity, pH, nutrient solution dissolved oxygen concentration, reflux water storage tank liquid level height, and concentrations of dissolved calcium ions, magnesium ions, potassium ions, and nitrate ions. An environmental field climate sensor matrix is ​​used to collect environmental field climate data, including air temperature, relative humidity, photosynthetically active radiation intensity, carbon dioxide concentration, saturated vapor pressure difference, crop surface leaf temperature, and root zone substrate temperature. An energy operation status sensor matrix is ​​used to collect energy operation data, including the remaining charge status of chemical energy storage batteries, the health degradation ratio of chemical energy storage batteries, the internal equivalent impedance value of chemical energy storage batteries, the instantaneous power generation of distributed photovoltaic units, and the current total electrical load of the system. Crop canopy image multispectral acquisition equipment is used to collect crop physiological phenotypic data, including the relative distribution of chlorophyll in crop leaves, leaf area index of the whole plant, root length, plant disease index, and fruit sugar content.

[0034] The specific logic for filtering, cleaning, and spatiotemporal alignment includes: By using the sliding unscented Kalman filter algorithm, the continuous analog electrical signals of environmental field climate data and energy operation data in each dimension are estimated, and noise and electromagnetic pulse noise in the sensor acquisition process are filtered out. It should be noted that the sliding unscented Kalman filter algorithm is a nonlinear filtering algorithm based on unscented transformation. It selects a set of deterministic sampling points near the state mean, called Sigma points, and directly performs state prediction and covariance update using the nonlinear system equations. Compared with traditional filtering, it has better nonlinear approximation capabilities. In this system, it is used to dynamically filter out outliers from sensor electromagnetic interference.

[0035] By using high-order Lagrange polynomial interpolation, brackish water ion field data with different acquisition frequencies are corrected and aligned to a uniform 1-hour clock step, and the equivalent value of brackish water ion field data at the standard control time is calculated. The irregularly distributed crop physiological phenotypic data are interpolated and reconstructed onto a unified greenhouse 3D grid node using the Kriging spatial interpolation algorithm. It should be noted that the Kriging spatial interpolation algorithm is a geostatistical interpolation method based on spatial correlation. It analyzes the spatial variogram structure of known measurement points to perform optimal unbiased estimation of unknown grid nodes. In this system, it is used to fuse irregularly distributed crop physiological sensor data onto a unified three-dimensional grid.

[0036] All multi-source physical state feature data that have been filtered, cleaned and spatiotemporally aligned are uniformly normalized and then one-dimensionally concatenated to obtain a state data vector.

[0037] Specifically, each filtered, cleaned, and spatiotemporally aligned data is normalized to the minimum and maximum values ​​and mapped to the [0,1] interval. The data is then concatenated into a unified state data vector in a fixed one-dimensional order. This state data vector eliminates the dimensional influence between physical quantities and is the only input data source for the calculation module.

[0038] The calculation module is used to extract state data vectors from the current moment to the past 24 hours and construct resource matrix, environment matrix and health matrix; The row vectors of the resource matrix are, in order, the desalination system water production efficiency, reverse osmosis membrane recovery rate, photovoltaic utilization rate, water and fertilizer recycling recovery rate, and root zone conductivity variation variance; The row vectors of the environmental matrix are, in order: air temperature fluctuation rate, air relative humidity fluctuation rate, photosynthetically active radiation intensity, saturated vapor pressure difference, weather forecast temperature deviation, and root zone dissolved oxygen fluctuation rate. The row vectors of the health matrix are, in order, the relative distribution of chlorophyll, the leaf area index of the whole plant, the root length, the fruit sugar content, and the plant disease index.

[0039] Specifically, the columns of the resource matrix, environment matrix, and health matrix represent data at each hour within a 24-hour period.

[0040] The data processing logic of the computing module includes: Max-min-extreme boundary normalization is performed on each row vector of the resource matrix, environment matrix, and health matrix to obtain dimensionless resource feature matrix, environment feature matrix, and health feature matrix; Principal component analysis is used to project each feature matrix onto the corresponding two-dimensional principal feature projection space, and two-dimensional discrete grid cells are divided in the two-dimensional principal feature projection space. The probability of sample points falling into each grid cell is statistically analyzed to obtain the discrete probability of the state distribution corresponding to each feature matrix. Specifically, the mean vectors along the column direction of the transposed resource feature matrix, environmental feature matrix, and health feature matrix are calculated respectively. The transposed feature matrices are then centered by zero mean, resulting in three centered feature matrices. The covariance matrix of each centered feature matrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix. The eigenvector columns corresponding to the two largest eigenvalues ​​are extracted, and a two-dimensional spatial projection transformation matrix is ​​constructed. Each centered feature matrix is ​​then projected onto the two-dimensional principal feature projection space composed of the corresponding two-dimensional spatial projection transformation matrix, and the projection coordinate matrix is ​​calculated. Within the obtained two-dimensional principal feature projection space, the extreme boundary of the planar coordinate point set in the horizontal and vertical directions is determined. The bounding box space region determined by the extreme boundary is uniformly divided into two-dimensional discrete grid cells of 3 rows and 3 columns. The number of sample points falling into each discrete grid cell is counted, and the probability of a sample point falling into each grid cell is calculated.

[0041] Based on Shannon's information entropy formula, the basic information entropy of the resource matrix, environment matrix, and health matrix are calculated respectively, and normalized by combining the historical maximum information entropy recorded in historical data to obtain normalized resource information entropy, normalized environment information entropy, and normalized health information entropy. Specifically, basic information entropy The mathematical expression is: ; in, Given the basic information entropy, k=1, 2, 3, These correspond to the resource matrix, environment matrix, and health matrix, respectively, representing the disorder of the corresponding feature matrix. 'a' is the row index of the two-dimensional discrete grid on the principal component projection plane, taking values ​​from 1 to 3 (positive integers). 'b' is the column index of the two-dimensional discrete grid on the principal component projection plane, taking values ​​from 1 to 3 (positive integers). The statistical probability that a historical sample point of the corresponding feature matrix falls into the grid cell of row a and column b; The probability term operation mapping function has the following piecewise logic: when hour, ; when hour, ; in, This is the natural logarithm operator, with the natural constant e as the base; The basic information entropy is compared with the historical maximum information entropy recorded in historical data. The logic is as follows: if the basic information entropy calculated at the current moment is... If the current base entropy is greater than the historical maximum entropy, then the historical maximum entropy will be updated to the current base entropy in the historical database. ; If the basic information entropy calculated at the current moment is If the value is less than or equal to the historical maximum information entropy, then the historical maximum information entropy value in the historical database will remain unchanged.

[0042] Dividing each basic information entropy by the final historical maximum information entropy yields the normalized resource information entropy, normalized environmental information entropy, and normalized health information entropy.

[0043] It should be noted that Shannon information entropy is a mathematical measure in information theory that measures the degree of uncertainty or disorder in a system. In this invention, by reducing the dimensionality of the business matrix and dividing it into discrete grids, the probability of sample points falling into the grids is statistically analyzed to calculate their information entropy values. The higher the entropy value of a certain dimension, the greater the degree of disorder and the severity of deviation from the target equilibrium state of that business characteristic indicator within the historical sliding window.

[0044] Calculate the proportion of each normalized information entropy in the sum of the three to obtain the dynamic allocation weights of resources, environment, and health.

[0045] In one embodiment of the present invention, the normalized information entropy represents the normalized uncertainty measure of each feature matrix. When the distribution of a certain business feature indicator fluctuates drastically or deviates from the equilibrium state, the calculated normalized information entropy increases, and the weight of that feature dimension also increases, thereby realizing the adaptive tilt of the system's adjustment computing power towards high-risk feature dimensions.

[0046] The multi-objective evaluation constraint equations include sub-objective functions for maximizing resource utilization, optimal environmental adaptation, maximizing plant health, maximizing system stability, and maximizing operational safety, and their mathematical expressions are as follows: ; , ; in, For multi-objective evaluation constraint equations, Candidate control vectors, For the sub-objective function of maximizing resource utilization, The optimal environmental adaptation sub-objective function. For the maximum plant health sub-objective function, The objective function is the sub-objective function for maximizing system stability. For the maximum operational safety sub-objective function, Dynamically assign weights to resources. Dynamically assign weights to the environment. Dynamically assign weights to health status. As the weight of the equipment's stable state, As meteorological disaster risk weights, That is , That is , They are respectively The final dynamic weights after normalization.

[0047] Specifically, the weight of the stable state of the equipment The mathematical expression is: ; in, For battery health; Meteorological disaster risk weight This represents the probability of rainfall in the current weather forecast. When the battery ages significantly (SOH decreases) or severe weather is approaching (probability of rainfall increases), the system automatically improves through normalization. or The proportion of [something] is used to limit drastic equipment movements and ensure the safety of the microgrid.

[0048] Candidate control vector This includes control variables such as desalination pump pressure, reflux ratio, irrigation EC, irrigation pH, LED supplemental lighting power, fan speed, and heating power.

[0049] The maximum resource utilization sub-objective function is used to calculate the product of the proportion of total water and fertilizer reuse in the total input raw materials and the photovoltaic power generation self-sufficiency rate; Specifically, the sub-objective function for maximizing resource utilization The mathematical expression is: ; in, To recover and reuse the total amount of water and fertilizer, For the newly prepared desalinated water volume, This refers to the amount of exogenous fertilizer input. This represents the real-time self-sufficiency rate of the local photovoltaic-storage microgrid.

[0050] The optimal environmental adaptation sub-objective function is used to calculate the deviation of the current air temperature and relative humidity from the suitable growth target values; Specifically, the optimal environment adaptation sub-objective function The mathematical expression is: ; in, The expected response temperature of the greenhouse air after executing the candidate control action u. The optimal temperature reference target value, The expected relative humidity of the greenhouse air after executing the candidate control action u. This is the optimal humidity reference target value.

[0051] The maximum plant health sub-objective function is used to calculate the deviation of the current leaf area index and relative chlorophyll distribution from the optimal reference value.

[0052] Specifically, the maximum plant health sub-objective function The mathematical expression is: ; in, and These represent the total leaf area index and relative chlorophyll content of the crop under the current conditions. and These are the optimal reference growth indicators for the corresponding growth stages of the crop.

[0053] The maximum system stability sub-objective function is used to apply action penalties to the single-step adjustment of the reverse osmosis system operating pressure and the single-step change of the energy storage battery discharge depth. Specifically, the maximum system stability sub-objective function The mathematical expression is: ; in, This represents the single-step change in operating pressure of the reverse osmosis system. The rated pressure limit, This represents the single-step change in the depth of discharge of the energy storage battery. To limit the maximum permissible single-step depth of discharge variation, the device life is extended by penalizing drastic changes in pressure and battery depth of discharge.

[0054] The maximum operational safety sub-objective function is used to calculate the operational safety protection level under external severe weather conditions based on the predicted rainfall probability and solar radiation intensity.

[0055] Specifically, the maximum operational safety sub-objective function The mathematical expression is: ; in, This represents the current probability of rainfall in the weather forecast. Normalized solar radiation intensity The higher the value, the higher the level of security protection of the system.

[0056] It should be noted that all of the above optimal reference values ​​were obtained through big data analysis.

[0057] At each control clock step, the decision module iteratively optimizes the multi-objective evaluation constraint equations in the constraint space using an adaptive multi-objective particle swarm optimization algorithm to find the optimal control vector and then issues it for execution. The control vectors include: desalination pump pressure, reflux ratio, irrigation EC and pH, LED supplemental lighting power, fan speed and heating power.

[0058] After the control action is executed, the self-learning module collects the first and second actual physical measurement values ​​of the sensor at the next moment, and calculates the mean square deviation (MSE) between the actual physical measurement value and the target set value. The first actual physical measurements include root zone matrix conductivity (EC), pH, greenhouse air temperature (T), relative humidity (RH), and reverse osmosis system operating pressure. The energy storage battery's state of charge (SOC) and the second actual physical measurement values ​​include the relative distribution of chlorophyll (SPAD) and the leaf area index (LAI).

[0059] When the calculated mean square deviation (MSE) exceeds the preset limit, the dynamic fine-tuning program is triggered to adjust the bias gain of the corresponding sub-target in the decision module online, thereby achieving closed-loop adaptive correction.

[0060] The self-learning module is equipped with a real-time feedback channel. After executing the control action at control cycle time t, it collects the first actual physical measurement value at the next control time t+1, and collects the actual measurement values ​​of slowly varying physiological quantities in the corresponding physiological sampling cycle. It calculates the mean square deviation (MSE) between the actual values ​​and the expected target values, and its mathematical expression is: ; in, This represents the actual measured value of the k-th physical parameter. Let M be the expected target value of the k-th physical parameter, and M be the number of parameter dimensions being evaluated.

[0061] If the mean square deviation (MSE) calculated corresponding to the first actual physical measurement value is greater than 0.05, i.e. the deviation is greater than 5%, the fast-change dynamic fine-tuning program is triggered to fine-tune the bias gain of the environmental and resource sub-objectives in the decision-making module online. If the mean square deviation calculated corresponding to the second actual physical measurement value exceeds the preset limit, the slow-varying dynamic fine-tuning program is triggered to correct the expected set value and weight gain of the plant health sub-target, ensuring that the system converges to the ideal control target quickly and smoothly.

[0062] The adjustment process of the dynamic fine-tuning program specifically includes: The self-learning module calculates the deviation based on the current control cycle and uses a proportional adaptive update method to correct the corresponding compensation parameters online. Its update expression is: ; ; in, For the previous control cycle The corresponding parameters to be corrected, For the next control cycle Updated parameters to be corrected The learning coefficient controls the update magnitude of the corresponding parameter; its value is a real number between 0 and 1. This refers to the control deviation of the corresponding physical quantity; Specifically, when the rapid dynamic fine-tuning procedure is triggered, the parameters to be corrected include temperature compensation correction coefficient, humidity compensation correction coefficient, and resource compensation correction coefficient. The initial value of the temperature compensation correction coefficient is the expected response temperature in the optimal environmental adaptation sub-objective function. The initial value of the humidity compensation correction coefficient is the expected response relative humidity in the optimal environmental adaptation sub-objective function, and the initial value of the resource compensation correction coefficient is the amount of newly prepared desalinated water. The updated resource compensation correction coefficient is used to correct the amount of newly produced desalinated water. Then it participates in the calculation of the maximum resource utilization sub-objective function.

[0063] When the slow-varying dynamic fine-tuning procedure is triggered, the parameters to be corrected include the leaf area index reference target value. Reference target value for relative distribution of chlorophyll content Dynamic weight allocation for health ; The updated parameters are used in the calculation of the multi-objective evaluation constraint equations for the next control cycle, enabling closed-loop online adaptive fine-tuning of the control parameters and ensuring that the system converges quickly and smoothly toward the ideal control objective.

[0064] For example, in a specific embodiment of the present invention, this embodiment includes a perception module acquiring and aligning multi-source data, a calculation module calculating normalized information entropy and dynamically allocating weights, a decision module constructing multi-objective evaluation constraint equations and optimizing and issuing control commands, and after executing the control action, a self-learning module collecting the first and second actual physical measurement values, calculating the mean square deviation (MSE), and fine-tuning the control parameters online through a proportional adaptive update formula. Configure a multi-dimensional sensor network to continuously acquire multi-source physical state characteristic data during greenhouse operation: Data collected using a brackish water ion field sensor matrix included the following parameters: nutrient solution reflux conductivity (2.2 mS / cm), pH (6.1), dissolved oxygen concentration (6.5 mg / L), reflux water storage tank level (1.2 m), and concentrations of dissolved calcium ions (180 mg / L), magnesium ions (60 mg / L), potassium ions (210 mg / L), and nitrate ions (450 mg / L).

[0065] Data collected via an environmental field climate sensor matrix included air temperature (25.2°C), relative humidity (68.5%), and photosynthetically active radiation intensity. The carbon dioxide concentration was 650 ppm, the saturated vapor pressure difference was 0.85 kPa, the temperature of the crop surface leaves was 24.5℃, and the temperature of the substrate in the root zone was 21.0℃.

[0066] Data collected through an energy operation status sensor matrix includes: State of Charge (SOC) of the chemical energy storage battery (SOC=82%), State of Health (SOH) of the chemical energy storage battery (SOH=88%), and the internal equivalent impedance value of the chemical energy storage battery. The instantaneous power generation of the distributed photovoltaic unit is 15.2kW and the current total electrical load of the system is 11.8kW.

[0067] Data collected using a crop canopy multispectral imaging device included the relative distribution of chlorophyll in crop leaves (SPAD=42.5), leaf area index (LAI=2.85), root length (18.2 cm), plant disease index (0.02), and fruit sugar content. .

[0068] A sliding unscented Kalman filter algorithm is used to perform state estimation on the continuous simulated electrical signals of environmental field climate data and energy operation data in each dimension, filtering out noise and electromagnetic pulse noise during the sensor acquisition process.

[0069] Using a high-order Lagrange polynomial interpolation method, brackish water ion field data with different acquisition frequencies are corrected and aligned to a uniform 1-hour clock step, and the equivalent value of brackish water ion field data at the standard control time is calculated.

[0070] The Kriging spatial interpolation algorithm is used to interpolate and reconstruct irregularly distributed crop physiological phenotypic data onto a unified greenhouse three-dimensional grid node.

[0071] The filtered, cleaned, and spatiotemporally aligned data are sequentially mapped to the [0,1] interval using max-min normalization, and then concatenated into a unified state data vector in a fixed one-dimensional order. This state data vector serves as the sole input data source for the computation module.

[0072] The calculation module extracts state data vectors from the current moment to the past 24 hours, constructing resource, environment, and health matrices. Column vectors represent data at each hour within the 24-hour period. Resource matrix (row vectors): desalination system water production efficiency, reverse osmosis membrane recovery rate, photovoltaic utilization rate, water and fertilizer recycling recovery rate, and root zone conductivity variation variance; Environmental matrix (row vectors): air temperature fluctuation rate, air relative humidity fluctuation rate, photosynthetically active radiation intensity, saturated vapor pressure difference, weather forecast temperature deviation, root zone dissolved oxygen fluctuation rate; Health matrix (row vectors): relative chlorophyll distribution content, whole plant leaf area index, root length, fruit sugar content, and plant disease index.

[0073] After transposing each feature matrix, zero mean centering is performed, the covariance matrix is ​​calculated and eigenvalue decomposition is performed, the eigenvector columns corresponding to the first two largest eigenvalues ​​are extracted to construct a two-dimensional spatial projection transformation matrix, which is then projected onto the two-dimensional principal feature projection space. Within the obtained two-dimensional principal feature projection space, the extreme value boundary bounding box is determined and uniformly divided into two-dimensional discrete grid cells of 3 rows × 3 columns. Count the number of times each discrete grid cell was filled in for the past 24 sample points, and calculate the probability of filling each cell. Let a ∈ {1, 2, 3} and b ∈ {1, 2, 3}. Taking the environmental feature matrix as an example, the statistical probability distribution of falling into each grid is shown in Table 1: Table 1: Statistical probability of the environmental feature matrix falling into a 3×3 two-dimensional discrete grid cell

[0074]

[0075] Substitute into Shannon's information entropy formula to calculate the basic information entropy. , ; The segmented mapping logic is as follows: when hour, ; when hour, .

[0076] Calculated environmental basic information entropy Similarly, the resource-based information entropy can be calculated. Health basic information entropy .

[0077] Compare the basic information entropy with the historical maximum information entropy Comparison and judgment, because ,Keep The value remains unchanged.

[0078] Dividing these values ​​yields the normalized information entropy: normalized resource information entropy 0.510, normalized environmental information entropy 0.802, and normalized health information entropy 0.387.

[0079] Calculate the proportion of each in the total to obtain the dynamic resource allocation weight. Dynamic allocation of environmental weights Health dynamic allocation weight .

[0080] Based on a battery health score (SOH) of 0.88, calculate the device's stable state weight. Read the rainfall probability from the weather forecast to obtain the meteorological disaster risk weight. .

[0081] according to Calculate the final dynamic weights. , , , , ; Constructing multi-objective evaluation constraint equations The candidate control vector u includes control variables such as desalination pump pressure, reflux ratio, irrigation EC, irrigation pH, LED supplemental lighting power, fan speed, and heating power.

[0082] The decision module iteratively optimizes the equation Fit(u) using an adaptive multi-objective particle swarm optimization algorithm, solves for it, and outputs the optimal control vector. Desalination pump pressure 1.85MPa, reflux ratio 75%, irrigation EC 2.1mS / cm, irrigation pH 5.8, LED supplemental lighting power. The fan speed is 60% and the heating power is 0W, and the command is sent to the actuator.

[0083] After the control action is issued, the first actual physical measurement value is evaluated: at the next control time t+1 (1 hour later), the self-learning module collects the first actual physical measurement value through the real-time feedback channel, including the root zone matrix conductivity EC, pH, greenhouse air temperature T, relative humidity RH, and reverse osmosis system operating pressure. The actual measured value of the state of charge (SOC) of the energy storage battery. With expected target value The comparative data are shown in Table 2: Table 2: Comparison of Actual Physical Measurement Values ​​and Expected Target Values ​​at Time t+1

[0084] Substituting into the mean square deviation formula, the MSE of the first actual physical measurement value is calculated as follows: MSE = 0.00613 + 0.00640 + 0.01417 + 0.00007 + 0.00657 + 0.000356 = 0.00561. Since the absolute value of the single deviation of the greenhouse air temperature T reaches -1.8℃, and the deviation is greater than 5%, the triggering condition for rapid dynamic fine-tuning is met.

[0085] Second Actual Physical Measurement Evaluation: At the corresponding 7-day physiological sampling period, second actual physical measurements were collected, including the relative distribution of chlorophyll (SPAD) and leaf area index (LAI). The measured actual LAI value was 2.10, the expected target value was 2.80, and the calculated MSE was greater than the preset limit, triggering the slow-varying dynamic fine-tuning procedure.

[0086] The self-learning module calculates the deviation e based on the current control cycle and performs online fine-tuning using a proportional adaptive update algorithm, triggering the fast-variable dynamic fine-tuning program. Temperature compensation correction coefficient: The initial value is the expected response temperature in the optimal environmental adaptation sub-objective function. =23.0℃. Taking the learning coefficient K=0.15, the deviation is 21.2-23.0=-1.8℃, and the temperature compensation correction coefficient for the next control cycle is updated to 23.0+0.15×(-1.8)=22.73℃; Humidity compensation correction coefficient: The initial value is the expected relative humidity RH = 65.0% in the optimal environmental adaptation sub-objective function. Taking the learning coefficient K = 0.10, the deviation is 70.2 - 65.0 = +5.2%. The updated humidity compensation correction coefficient for the next control cycle is 65.0 + 0.10 × (+5.2) = 65.52%. Resource compensation correction coefficient: The initial value is the amount of newly produced desalinated water in the maximum resource utilization sub-objective function. The resource learning coefficient K=0.08 was selected, and the deviation was controlled by the actual amount of freshly produced desalinated water. The resource compensation correction factor for the next control cycle is updated and calculated as follows: .

[0087] Trigger the slow-change dynamic fine-tuning procedure: Leaf area index reference target value With a learning coefficient K=0.10 and a bias of 2.10-2.80=-0.70, update the reference target value of leaf area index for the next control period. The value is 2.80 + 0.10 × (-0.70) = 2.73; Reference target value for relative chlorophyll content With a learning coefficient K=0.10 and a bias of 38.0-42.5=-4.5, update the reference target value for the relative distribution of chlorophyll content for the next control period. The result is 42.5 + 0.10 × (-4.5) = 42.05; Health dynamic allocation weight The original value was 0.228. The learning coefficient K was set to 0.20. The health dynamic allocation weights for the next control cycle were then updated based on the physiological deviation amplitude. for ; The self-learning module sends parameter correction signals to the decision-making module, substituting the updated temperature compensation correction coefficient of 22.73℃ and humidity compensation correction coefficient of 65.52% into the optimal environmental adaptation sub-objective function. ; Substitute the updated resource compensation correction coefficient into the maximum resource utilization sub-objective function. ; The updated leaf area index will be used as a reference target value. Reference target value for relative distribution of chlorophyll content Substitute into the maximum plant health sub-objective function ; Updated health dynamic allocation weights Re-entering the second normalization process, automatically boosting Weight ratios in multi-objective evaluation constraint equations.

[0088] The updated parameters participate in the optimization calculation of the multi-objective evaluation constraint equation in the next control cycle, driving the actuator to increase the heating power and LED supplementary light intensity, so that the actual response physical quantity quickly converges to the desired target, realizing high-precision and robust closed-loop adaptive correction of brackish water desalination and hydroponic control with low computing power overhead.

[0089] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0090] This system introduces information entropy from the business matrix to enhance the rationality and stability of system decisions. It continuously collects system state data through a multi-dimensional sensor network and uses a sensing module to perform filtering, cleaning, and spatiotemporal alignment processing to construct a unified state data vector. By performing principal component analysis and discrete grid probability statistics on the resource, environmental, and health data matrices within a historical sliding window, it calculates the normalized information entropy of each dimension, which serves as the dynamic weight allocation coefficient, achieving a seamless mapping between algorithmic indicators and actual agricultural business semantics. Adaptive multi-objective particle swarm optimization is used to solve for and distribute the optimal control vector. Through a closed-loop feedback self-learning channel, the actual response deviation of the physical system is detected in real time. When the deviation exceeds a preset limit, the control gain and target weight parameters are fine-tuned online, achieving adaptive closed-loop control throughout the entire lifecycle.

[0091] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0092] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A closed-loop AI-based precision water and fertilizer recycling hydroponic treatment system for brackish water desalination, characterized in that, It includes a perception module, a computing module, a decision-making module, and a self-learning module; The sensing module is configured with a multi-dimensional sensor network to acquire multi-source physical state feature data during the operation of the greenhouse, and to filter, clean and align the data in time and space to output a state data vector. The calculation module is used to construct a resource matrix, an environment matrix, and a health matrix based on the state data vector. After normalization, it is used to reduce the dimensionality based on the principal component analysis algorithm and calculate the discrete probability of the state distribution. The normalized information entropy is calculated by substituting it into the Shannon information entropy formula. Based on the normalized information entropy, the dynamic allocation weights of resources, environment, and health are obtained. The decision-making module is used to construct a multi-objective evaluation constraint equation, including resource utilization, environmental adaptability, plant health, system stability and operational safety, based on the dynamic allocation weights of resources, environment, and health, as well as the current state data vector and meteorological forecast data. It then optimizes and outputs the optimal control command and issues it for execution. The self-learning module is used to collect actual environmental response data after the execution of instructions, calculate the deviation statistics between the actual environmental response data and the expected target setting value, and send a parameter correction signal to the decision module when the deviation statistics are greater than the preset limit, so as to fine-tune the target weights and control gain parameters in the multi-objective evaluation constraint equation online.

2. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 1, characterized in that: The multi-dimensional sensor network includes a brackish water ion field sensor matrix, an environmental field climate sensor matrix, an energy operation status sensor matrix, and a crop canopy image multispectral acquisition device. The multi-source physical state characteristic data includes brackish water ion field data, environmental field climate data, energy operation data, and crop physiological phenotypic data; The multi-source physical state feature data is packaged and encapsulated.

3. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 2, characterized in that: The specific logic for filtering, cleaning, and spatiotemporal alignment includes: By using the sliding unscented Kalman filter algorithm, the continuous analog electrical signals of environmental field climate data and energy operation data in each dimension are estimated, and noise and electromagnetic pulse noise in the sensor acquisition process are filtered out. By using high-order Lagrange polynomial interpolation, brackish water ion field data with different acquisition frequencies are corrected and aligned to a uniform 1-hour clock step, and the equivalent value of brackish water ion field data at the standard control time is calculated. The irregularly distributed crop physiological phenotypic data are interpolated and reconstructed onto a unified greenhouse 3D grid node using the Kriging spatial interpolation algorithm. All multi-source physical state feature data that have been filtered, cleaned and spatiotemporally aligned are uniformly normalized and then one-dimensionally concatenated to obtain a state data vector.

4. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 1, characterized in that: The calculation module is used to extract state data vectors from the current moment to the past 24 hours and construct resource matrix, environment matrix and health matrix; The row vectors of the resource matrix are, in order, the desalination system water production efficiency, reverse osmosis membrane recovery rate, photovoltaic utilization rate, water and fertilizer recycling recovery rate, and root zone conductivity variation variance. The row vectors of the environmental matrix are, in order: air temperature fluctuation rate, air relative humidity fluctuation rate, photosynthetically active radiation intensity, saturated vapor pressure difference, weather forecast temperature deviation, and root zone dissolved oxygen fluctuation rate. The row vectors of the health matrix are, in order, the relative distribution of chlorophyll, the leaf area index of the whole plant, the root length, the fruit sugar content, and the plant disease index.

5. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 4, characterized in that: The data processing logic of the computing module includes: The maximum-minimum-extreme-boundary normalization is performed on each row vector of the resource matrix, environment matrix, and health matrix to obtain dimensionless resource feature matrix, environment feature matrix, and health feature matrix. Principal component analysis is used to project each feature matrix onto the corresponding two-dimensional principal feature projection space, and two-dimensional discrete grid cells are divided in the two-dimensional principal feature projection space. The probability of sample points falling into each grid cell is statistically analyzed to obtain the discrete probability of the state distribution corresponding to each feature matrix. Based on Shannon's information entropy formula, the basic information entropy of the resource matrix, environment matrix, and health matrix are calculated respectively, and normalized by combining the historical maximum information entropy recorded in historical data to obtain normalized resource information entropy, normalized environment information entropy, and normalized health information entropy. Calculate the proportion of each normalized information entropy in the sum of the three to obtain the dynamic allocation weights of resources, environment, and health.

6. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 1, characterized in that: The multi-objective evaluation constraint equations include sub-objective functions for maximizing resource utilization, optimal environmental adaptation, maximizing plant health, maximizing system stability, and maximizing operational safety. Their mathematical expressions are as follows: ； ， ； in, For multi-objective evaluation constraint equations, Candidate control vectors, For the sub-objective function of maximizing resource utilization, The optimal environmental adaptation sub-objective function. For the maximum plant health sub-objective function, The objective function is the sub-objective function for maximizing system stability. For the maximum operational safety sub-objective function, Dynamically assign weights to resources. Dynamically assign weights to the environment. Dynamically assign weights to health status. As the weight of the equipment's stable state, As meteorological disaster risk weights, That is , That is , They are respectively The final dynamic weights after normalization.

7. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 6, characterized in that: The maximum resource utilization sub-objective function is used to calculate the product of the proportion of total water and fertilizer reuse in the total input raw materials and the photovoltaic power generation self-sufficiency rate. The optimal environmental adaptation sub-objective function is used to calculate the deviation of the current air temperature and relative humidity from the suitable growth target value; The maximum plant health sub-objective function is used to calculate the deviation of the current leaf area index and the relative chlorophyll distribution from the optimal reference value.

8. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 7, characterized in that: The maximum system stability sub-objective function is used to apply action penalties to the single-step adjustment of the reverse osmosis system operating pressure and the single-step change in the discharge depth of the energy storage battery. The maximum operational safety sub-objective function is used to calculate the operational safety protection level under external severe weather conditions based on the rainfall probability and solar radiation intensity predicted by the weather forecast.

9. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 1, characterized in that: At each control clock step, the decision module iteratively optimizes the multi-objective evaluation constraint equations in the constraint space using an adaptive multi-objective particle swarm optimization algorithm to find the optimal control vector and then issues it for execution. The control vectors include: desalination pump pressure, reflux ratio, irrigation EC and pH, LED supplemental lighting power, fan speed and heating power.

10. The brackish water desalination closed-loop AI precision water and fertilizer circulation hydroponic treatment system according to claim 9, characterized in that: After the control action is executed, the self-learning module collects the first and second actual physical measurement values ​​of the sensor at the next moment and calculates the mean square deviation (MSE) between the actual physical measurement value and the target set value. When the calculated mean square deviation (MSE) exceeds the preset limit, a dynamic fine-tuning procedure is triggered to adjust the bias gain of the corresponding sub-target in the decision module online, thereby achieving closed-loop adaptive correction.

Images