Facility greenhouse water and fertilizer-environment monitoring system based on physical information neural network

Abstract

The invention discloses a facility greenhouse water and fertilizer-environment monitoring system based on a physical information neural network. Belongs to the technical field of facility agriculture intelligent control and particularly relates to the technical field of environmental monitoring based on a physical information neural network. The problem that a previous facility greenhouse water and fertilizer-environment monitoring method cannot truly reflect the coupling relation between crops and the environment is solved. The system comprises a sensor module which obtains facility greenhouse environment data under each time-space coordinate in real time; the multi-influence-factor coupling construction module is used for carrying out joint modeling on factors influencing the water, fertilizer and environment of the facility greenhouse to obtain a multi-influence-factor coupling equation; and the physical information neural network prediction module is used for designing a physical information neural network architecture, defining a loss function of a physical information neural network based on the multi-influence-factor coupling model, and obtaining the predicted water and fertilizer-environment state of the facility greenhouse through the physical information neural network.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent control in protected agriculture, and specifically relates to the technical field of environmental monitoring based on a physics-informed neural network. Background Art

[0002] With the continuous development of modern agriculture, protected structures (such as greenhouses and high-wire cultivation systems), as key technologies to improve crop production efficiency and stability, have received increasing attention. Protected structures can dynamically optimize crop growth conditions by precisely controlling environmental factors such as light, temperature, humidity, and gas, thereby increasing crop yield, quality, and stress resistance. However, in the environment of protected structures, the interaction between multiple factors such as water and fertilizer supply, light intensity, and CO2 concentration is complex and dynamic, which poses challenges for precise water and fertilizer regulation.

[0003] Traditional control methods for protected structures often regulate different environmental factors (such as water and fertilizer supply and climate control) separately, resulting in the failure to effectively consider the interaction between different influencing factors. Especially under high-temperature conditions, the transpiration of crops intensifies, increasing the consumption of substrate moisture, and high light intensity further accelerates water evaporation, leading to an imbalance in the supply of water and nutrients. This not only fails to truly reflect the coupling relationship between crops and the environment but also makes it difficult for traditional control methods to dynamically respond to changes in water and fertilizer requirements during the crop growth process. Especially during the critical period of crop growth, irrigation strategies often cannot be adjusted in a timely manner, affecting crop growth and yield.

[0004] In addition, the substrates used in protected structures (such as coconut coir, rock wool, etc.) have complex pore structures and hydraulic characteristics, which makes it difficult for traditional water movement models based on the Richards equation to accurately describe the dynamic water characteristics of the substrates. Although traditional water movement models (such as the Van Genuchten model) can describe the water behavior of substrates to a certain extent, their parameters are usually static and difficult to calibrate dynamically, resulting in large errors in model prediction during actual operation and being unable to accurately reflect the responses of different substrates to environmental changes.

[0005] Furthermore, traditional control methods often make predictions based on historical data or static rules. However, due to insufficient training data, pure data-driven models are prone to prediction lags in short cycles, resulting in the nutrient supply being unable to meet the rapid growth needs of crops, thereby affecting crop yield and quality. Summary of the Invention

[0006] In order to solve the problem that the previous water and fertilizer - environmental monitoring methods for protected structures have large prediction errors for substrate status and cannot truly reflect the coupling relationship between crops and the environment, the present invention provides a water and fertilizer - environmental monitoring system for protected structures based on a physics-informed neural network. The system includes:

[0007] Sensor module: used to obtain the environmental data of the facility greenhouse at each spatio-temporal coordinate in real time;

[0008] Multi-factor coupling construction module: used to jointly model the factors affecting the water-fertilizer-environment of the facility greenhouse to obtain a multi-factor coupling equation;

[0009] The multi-factor coupling model includes a substrate moisture coupling transport equation, a substrate nutrient coupling transport equation, and a light-temperature response growth equation for facility crops;

[0010] Physics-informed neural network prediction module: used to design the architecture of the physics-informed neural network and define the loss function of the physics-informed neural network based on the multi-factor coupling model;

[0011] Used to obtain the predicted water-fertilizer-environment state of the facility greenhouse through the physics-informed neural network.

[0012] Furthermore, the spatio-temporal coordinate is (z, t), where z represents the soil depth and t represents time; the facility greenhouse environmental data includes photosynthetically active radiation, air temperature, and CO2 concentration.

[0013] Furthermore, the substrate moisture coupling transport equation is specifically: where θ represents the substrate moisture content, K(θ, T) represents the temperature-dependent hydraulic conductivity, ψ represents the substrate potential, α represents the inclination angle of the facility greenhouse cultivation tank, S root (PAR, C) represents the root absorption function, and the root absorption function is determined by the photosynthetically active radiation PAR and the nutrient solution concentration C.

[0014] Furthermore, the substrate nutrient coupling transport equation is specifically: where D(T) represents the diffusion coefficient at temperature T, v(θ) represents the water flow velocity at the substrate moisture content θ, k uptake (PAR, T root ) represents the light-temperature regulated root absorption coefficient, and the light-temperature regulated root absorption coefficient is determined by the photosynthetically active radiation PAR and the root temperature T root decide.

[0015] Furthermore, the light-temperature response growth equation for facility crops is specifically: where W represents the crop biomass, ∈ LUE represents the light energy utilization efficiency, f(T leaf ) represents the temperature response function, which is determined by the leaf temperature T leaf decide, g(C xylem ) represents the xylem nutrient transport efficiency, which is determined by the nutrient concentration C in the xylem xylemDecision, R d (T night ) represents the dark respiration rate, which is determined by the night temperature T night .

[0016] Furthermore, the physics-informed neural network architecture includes an input layer, a hidden layer, and an output layer, specifically:

[0017] Input layer: Inputs spatio-temporal coordinates and facility greenhouse environment data;

[0018] Hidden layer: Adopts 6 hidden layers, with 256 nodes in each layer, and the activation function is the Swish activation function;

[0019] Output layer: Outputs the predicted water-fertilizer - environment state of the facility greenhouse, including the predicted substrate moisture content at the current spatio-temporal coordinates nutrient solution concentration and crop biomass

[0020] Furthermore, the loss function of the physics-informed neural network is specifically:

[0021]

[0022] where λ1, λ2, and λ3 are the coefficients of each term in the loss function; MSE data represents the data fitting term; MSE physics represents the physical constraint term; MSE boundary represents the boundary condition term.

[0023] Furthermore,

[0024] where N represents the total number of measured data points, and each measured data point corresponds to a specific spatio-temporal coordinate, represents the substrate moisture content of the data point predicted by the physics-informed neural network, represents the nutrient solution concentration of the data point predicted by the physics-informed neural network, represents the crop biomass of the data point predicted by the physics-informed neural network.

[0025] Furthermore,

[0026]

[0027] where N p represents the number of collocation points, and each collocation point corresponds to a specific spatio-temporal coordinate, represents the residual of the substrate moisture coupled transport equation, represents the residual of the substrate nutrient coupled transport equation, represents the residual of the facility crop light-temperature response growth equation.

[0028] Furthermore, N b represents the number of boundary data points, and each boundary data point corresponds to a specific spatio-temporal coordinate. represents the result after replacing the substrate moisture content θ with the predicted substrate moisture content at the current spatio-temporal coordinate in the calculation formula of K(θ,T). The result after represents the predicted substrate potential, E p (t k ) is the measured transpiration rate of the sap flow meter at time t. k The beneficial effects of the system of the present invention are as follows:

[0029] The beneficial effects of the system of the present invention are as follows:

[0030] (1) Based on the physics-informed neural network (PINN), complex influencing factors (including substrate moisture-nutrient coupling, crop growth models, and environmental regulation mechanisms) are effectively coupled in a unified framework. The loss function and network architecture are designed so that the neural network can not only efficiently learn the potential patterns in the data but also follow the known physical laws, thereby ensuring that the system can accurately predict multiple factors.

[0031] (2) In the facility greenhouse environment, the mutual influence among multiple factors such as water and fertilizer supply, light intensity, and CO2 concentration is complex and dynamic. The parameters of traditional models (such as the Van Genuchten model) that reflect the transport of substrate moisture and nutrients are fixed and cannot dynamically adapt to environmental changes. The system of the present invention constructs a multi-influencing factor coupling model, and introduces temperature-dependent hydraulic conductivity and light-temperature-regulated root absorption coefficients into the substrate moisture coupling transport equation and the substrate nutrient coupling transport equation, so as to accurately reflect the response of the substrate in the real environment. Traditional crop light-temperature response growth models (such as empirical formulas) ignore the dynamic influence of the nutrient transport efficiency of the woody part. The system of the present invention constructs a light-temperature response growth equation for facility crops, and comprehensively captures the non-linear characteristics of crop growth by introducing a light-temperature response function and a nutrient efficiency function. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a schematic diagram of the system according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0034] Example 1

[0035] This example provides a water-fertilizer-environment monitoring system for facility greenhouses based on a physics-informed neural network, as Figure 1 shown. The system includes:

[0036] Sensor module: used to obtain the environmental data of the facility greenhouse at each spatio-temporal coordinate in real time;

[0037] Multi-factor coupling construction module: used to jointly model the factors affecting the water-fertilizer-environment of the facility greenhouse to obtain a multi-factor coupling equation;

[0038] The multi-factor coupling model includes a substrate moisture coupling transport equation, a substrate nutrient coupling transport equation, and a light-temperature response growth equation for facility crops;

[0039] Physics-informed neural network prediction module: used to design the architecture of the physics-informed neural network and define the loss function of the physics-informed neural network based on the multi-factor coupling model;

[0040] Used to obtain the predicted water-fertilizer-environment state of the facility greenhouse through the physics-informed neural network.

[0041] Example 2

[0042] This example further limits Example 1. In the sensor module, the photosynthetically active radiation PAR, air temperature T air and CO2 concentration are obtained through a canopy sensor; the substrate moisture content θ is obtained through a TDR moisture sensor, and the nutrient solution concentration C is obtained through an EC probe. In the sensor module, other crop parameters can also be obtained according to needs, and the transpiration rate E p (t) is measured using a sap flow meter and the crop biomass W is obtained by biomass weighing or NDVI inversion.

[0043] The spatio-temporal coordinate is (z, t), where z represents the soil depth and t represents the time.

[0044] In specific applications, the spatio-temporal coordinate (z, t) can also be extended to (x, y, z, t), where x represents the lateral coordinate and y represents the longitudinal coordinate. Extending the spatio-temporal coordinate can make the prediction more accurate.

[0045] Example 3

[0046] This example further limits Example 1. The substrate moisture coupling transport equation is specifically: where θ represents the substrate moisture content, K(θ, T) represents the temperature-dependent hydraulic conductivity, ψ represents the substrate potential, α represents the inclination angle of the facility greenhouse cultivation tank, S root(PAR, C) represents the root absorption function, which is determined by the photosynthetically active radiation PAR and the nutrient solution concentration C.

[0047] is the temperature-dependent hydraulic conductivity, α is the inclination angle of the cultivation tank, K s represents a measure of the water conductivity of the substrate in the fully saturated state, θ r represents the minimum water content in the substrate that cannot be utilized by plants, θ s represents the water content when the substrate is fully saturated, E a represents the temperature sensitivity of water movement, which indicates the ability of temperature to affect water movement. R represents the gas constant, whose value is approximately 8.314 J / (mol·K), and is used to calculate the exponential term of the temperature effect. T represents the current temperature, and T0 represents the reference temperature used for temperature comparison.

[0048] The specific substrate nutrient coupled transport equation is as follows: Among them, D(T) represents the diffusion coefficient at temperature T, v(θ) represents the water flow velocity at the substrate water content θ, k uptake (PAR, T root ) represents the root absorption coefficient regulated by light and temperature, which is determined by the photosynthetically active radiation PAR and the root temperature T root determined.

[0049] k uptake = k0·sigmoid(PAR - 800)·exp(-|T root - 25| / 10);

[0050] The specific light and temperature response growth equation of protected crops is as follows: Among them, W represents the crop biomass, ∈ LUE represents the light energy utilization efficiency, f(T leaf ) represents the temperature response function, which is determined by the leaf temperature T leaf determined, g(C xylem ) represents the xylem nutrient transport efficiency, which is determined by the nutrient concentration C in the xylem xylem determined, R d (T night ) represents the dark respiration rate, which is determined by the night temperature T night determined.

[0051] f(T leaf ) = exp(-|T leaf - 25| / 8), g(C xylem ) = tanh(C xylem / 2).

[0052] Example 4 This embodiment further defines Embodiment 1. The physical information neural network architecture includes an input layer, a hidden layer, and an output layer, specifically: Input layer: Inputs spatio-temporal coordinates and facility greenhouse environmental data; Hidden layer: Adopts 6 hidden layers, with 256 nodes in each layer, and the activation function is the Swish activation function; Output layer: Outputs the predicted water-fertilizer - environmental status of the facility greenhouse, including the predicted substrate moisture content at the current spatio-temporal coordinates , nutrient solution concentration and crop biomass .

[0057] Analyze the trends and changes of these parameters to determine crop requirements and growth status, and perform corresponding regulation according to different crops, thereby helping users to dynamically regulate the water-fertilizer - environment of the facility greenhouse.

[0058] The loss function of the physical information neural network is specifically:

[0059]

[0060] where λ1, λ2, and λ3 are the coefficients of each term in the loss function; MSE data represents the data fitting term; MSE physics represents the physical constraint term; MSE boundary represents the boundary condition term.

[0061]

[0062] where N represents the total number of measured data points, and each measured data point corresponds to a specific spatio-temporal coordinate, represents the substrate moisture content of the data point predicted by the physical information neural network, represents the nutrient solution concentration of the data point predicted by the physical information neural network, represents the crop biomass of the data point predicted by the physical information neural network.

[0063]

[0064] where N p represents the number of collocation points. Collocation points are a term used in numerical simulations and machine learning models. These points are used to ensure that the model output conforms to predetermined physical laws or equations; each collocation point corresponds to a spatio-temporal coordinate, represents the residual of the substrate moisture coupled transport equation, Represents the residual of the coupled transport equation of substrate nutrients, represents the residual of the light and temperature response growth equation of protected crops.

[0065] N b represents the number of boundary data points, and each boundary data point corresponds to a spatio-temporal coordinate, represents the result after replacing the substrate water content θ with the predicted substrate water content at the current spatio-temporal coordinate in the calculation formula of K(θ,T) after that, represents the predicted substrate potential, E p (t k ) is the transpiration rate measured by the sap flow meter at time t k .

[0066] Example 5,

[0067] This example further limits Example 1 and further introduces the training of the physical information neural network and the calculation of the weight coefficients.

[0068] Initialization: Randomly initialize the network weights, set the initial learning rate η = 10 -3 , weight coefficients λ1 = 1.0, λ2 = 0.1, λ3 = 0.5.

[0069] Forward propagation: Calculate the network output and the loss function

[0070] Backward propagation: Calculate the gradient of the loss function with respect to the network weights and update the weights:

[0071] Dynamically adjust the weights: Dynamically adjust according to the ratio of the physical residual to the data error:

[0072] where, represents and 's variance, reflecting the degree to which the model deviates from the physical laws;

[0073] Var(θ obs ,C obs ,W obs ) represents the variance of the substrate water content θ, nutrient solution concentration C, and crop biomass W obtained by the sensor module, reflecting the degree of data noise or sensor abnormality.

[0074] where CV represents the coefficient of variation of the transpiration rate, E p (t) represents the transpiration rate.

[0075] Among them, θ(z = 0, t) represents the surface substrate water content, and |E p (t)|² represents the L2 norm of the transpiration rate E p (t), which characterizes the transpiration intensity.

[0076] |θ(z = 0, t)|² represents the L2 norm of the surface substrate water content θ(z = 0, t), reflecting the surface water status, ∈ = 10 -5 , and its function is to prevent and eliminate zero.

[0077] Iterative optimization: Repeat the steps until the loss function converges or the number of iterations reaches.

[0078] Prediction accuracy: The RMSE of substrate water content prediction is 0.018 m 3 / m 3 , and the RMSE of nutrient concentration prediction is 0.23 mg / L.

[0079] Expected control effect: The water and fertilizer utilization efficiency is increased by 15%.

Claims

1. A water and fertilizer - environment monitoring system for facility greenhouses based on a physics - informed neural network, characterized in that, The system includes: A sensor module: used to obtain the environmental data of the facility greenhouse at each spatio-temporal coordinate in real time; A multi-influence-factor coupling construction module: used to jointly model the factors affecting the water-fertilizer-environment of the facility greenhouse to obtain a multi-influence-factor coupling equation; The multi-influence-factor coupling model includes a matrix moisture coupling transport equation, a matrix nutrient coupling transport equation, and a light-temperature response growth equation for facility crops; A physics-informed neural network prediction module: used to design the architecture of the physics-informed neural network and define the loss function of the physics-informed neural network based on the multi-influence-factor coupling model; Used to obtain the predicted water-fertilizer-environment state of the facility greenhouse through the physics-informed neural network.

2. The water and fertilizer - environment monitoring system for facility greenhouses based on the physics - informed neural network according to claim 1, wherein, The spatio-temporal coordinate is (z, t), where z represents the soil depth and t represents the time; the facility greenhouse environmental data includes photosynthetically active radiation, air temperature, and CO2 concentration.

3. The facility shed water and fertilizer - environment monitoring system based on the physics - informed neural network according to claim 2, characterized in that, The specific matrix moisture coupled migration equation is as follows: where θ represents the matrix moisture content, K(θ,T) represents the temperature-dependent hydraulic conductivity, ψ represents the matrix potential, α represents the inclination angle of the cultivation trough in the greenhouse, and S root (PAR,C) represents the root absorption function, which is determined by the photosynthetically active radiation PAR and the nutrient solution concentration C.

4. The facility shed water and fertilizer - environment monitoring system based on the physics - informed neural network according to claim 3, characterized in that, The specific matrix nutrient coupled migration equation is as follows: where D(T) represents the diffusion coefficient at temperature T, v(θ) represents the water flow velocity at matrix water content θ, and k uptake (PAR, T root ) represents the root absorption coefficient regulated by light and temperature, and the root absorption coefficient regulated by light and temperature is determined by the photosynthetically active radiation PAR and the root temperature T root .

5. The water and fertilizer - environment monitoring system for facility shed based on the physics - informed neural network according to claim 4, characterized in that, The light and temperature response growth equation of the protected crop is specifically as follows: where W represents the crop biomass, ∈ LUE represents the light energy utilization efficiency, f(T leaf ) represents the temperature response function, which is determined by the leaf temperature T leaf , g(C xylem ) represents the xylem nutrient transport efficiency, which is determined by the nutrient concentration C xulem in the xylem, R d (T night ) represents the dark respiration rate, which is determined by the night temperature T night .

6. The water and fertilizer - environment monitoring system for facility greenhouses based on the physics - informed neural network according to claim 5, wherein The architecture of the physics-informed neural network includes an input layer, a hidden layer, and an output layer, specifically: Input layer: Input the spatio-temporal coordinates and the environmental data of the facility greenhouse; Hidden layer: Adopt 6 hidden layers, with 256 nodes in each layer, and the activation function is the Swish activation function; Output layer: Output the predicted water-fertilizer-environment status of the facility greenhouse, including the predicted substrate moisture content at the current spatio-temporal coordinates Nutrient solution concentration and crop biomass 7. The water and fertilizer - environment monitoring system for facility greenhouses based on the physics - informed neural network according to claim 6, characterized in that, The loss function of the physics-informed neural network is specifically: Among them, λ1, λ2, and λ3 are the coefficients of each term in the loss function; MSE data represents the data fitting term; MSE physics represents the physical constraint term; MSE boundary represents the boundary condition term.

8. The water-fertilizer-environment monitoring system for facility greenhouses based on a physics-informed neural network according to claim 7, characterized in that Among them, N represents the total number of measured data points, and each measured data point corresponds to a specific spatio-temporal coordinate. represents the substrate moisture content of the data point predicted by the physics-informed neural network. represents the nutrient solution concentration of the data point predicted by the physics-informed neural network. represents the crop biomass of the data point predicted by the physics-informed neural network.

9. The water-fertilizer-environment monitoring system for facility greenhouses based on a physics-informed neural network according to claim 8, characterized in that Among them, N p represents the number of configuration points, and each configuration point corresponds to a specific spatio-temporal coordinate, represents the residual of the matrix moisture coupled transport equation, represents the residual of the matrix nutrient coupled transport equation, represents the residual of the light and temperature response growth equation of protected crops.

10. The water-fertilizer-environment monitoring system for facility greenhouses based on a physics-informed neural network according to claim 9, characterized in that N b represents the number of boundary data points, and each boundary data point corresponds to a specific spatio-temporal coordinate. represents the result after replacing the substrate water content θ with the predicted substrate water content at the current spatio-temporal coordinate in the calculation formula of K(θ,T). The result after replacement. represents the predicted substrate potential, E p (t k ) is the measured transpiration rate of the sap flow meter at time t. k ​

Images