Greenhouse temperature field real-time prediction and regulation system based on physical information neural network

Abstract

The invention discloses a greenhouse temperature field real-time prediction and regulation system based on a physical information neural network. Relates to the technical field of intelligent agriculture and environment control, in particular to a greenhouse temperature field three-dimensional real-time prediction method based on a physical information neural network and the technical field of energy-saving regulation and control. According to the method, the minimum power of the heating equipment and the ventilation equipment is solved by combining the greenhouse temperature field prediction model and the multi-objective optimization model, and energy consumption is reduced while it is guaranteed that temperature field distribution meets crop growth requirements. The system comprises a data acquisition module used for acquiring space-time coordinates of a to-be-predicted position; the state prediction module is used for constructing a greenhouse temperature field prediction model and total loss; the processor is used for processing space-time coordinates of a to-be-predicted position and outputting a temperature field of the to-be-predicted position; the multi-objective optimization module is used for minimizing the power of the ventilation equipment and the heating equipment; and the execution module regulates and controls the greenhouse temperature field.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent agriculture and environmental control, and particularly to a three-dimensional real-time prediction method for greenhouse temperature field based on physical information neural network and an energy-saving regulation technology field. Background Technique

[0002] As a highly controlled agricultural production environment, the greenhouse plays an important role in promoting crop growth, yield and quality. The temperature field distribution inside the greenhouse directly affects physiological processes such as photosynthesis, respiration and water transpiration of crops. Therefore, precise temperature regulation is crucial in modern greenhouse agriculture. However, in practical applications, the monitoring and regulation of greenhouse temperature face many technical challenges.

[0003] Traditional greenhouse temperature field monitoring methods rely on densely deployed sensor networks. Although they can provide real-time data to a certain extent, this method has significant disadvantages. Traditional sensor networks have limitations in spatial resolution and cannot accurately capture the minute changes in the greenhouse temperature field. Especially in a greenhouse environment with complex space and irregular layout, the sparsity of sensor data further reduces the accuracy of temperature field prediction.

[0004] Among the existing greenhouse temperature field prediction methods, data-driven deep learning models (such as long short-term memory network LSTM, convolutional neural network CNN, etc.) can, to a certain extent, predict temperature trends and changes. However, these methods usually only model based on historical data and lack physical consistency.

[0005] The temperature field simulation based on pure physical models (such as the finite element method) can more accurately describe the heat conduction process, but its computational cost is high. Especially in the application scenario of real-time prediction, the calculation speed of the finite element method cannot meet the demand for real-time update. In the greenhouse environment, the change of temperature is affected by various factors such as light, crop transpiration, wind speed and greenhouse building structure, and it is required to be able to respond dynamically and quickly to these changes. In such a high-complexity scenario, it is often difficult for the physical model-based prediction method to achieve fast and accurate temperature prediction.

[0006] Currently, effectively combining complex physical equations (such as heat conduction, convection, radiation equations, etc.) with the computational framework of deep learning to ensure that the model can not only learn data-driven laws but also maintain physical consistency is still a technical problem. The introduction of physical equations will make the training process more complex, and it is difficult to balance the relationship between physical constraints and data-driven learning.

[0007] Therefore, it is necessary to design a suitable loss function to ensure that physical laws can be correctly embedded in the training process and will not overly restrict the learning ability of the neural network. Summary of the Invention

[0008] Aiming at the problems of traditional aquaculture environment temperature field regulation relying on empirical models, low computational efficiency, and insufficient accuracy in multi-field coupling modeling, the present invention proposes a real-time prediction and regulation system for greenhouse temperature fields based on physics-informed neural networks. By combining the greenhouse temperature field prediction model and the multi-objective optimization model, the minimum power of heating equipment and ventilation equipment is solved, and while ensuring that the temperature field distribution meets the growth requirements of crops, energy consumption is reduced.

[0009] The system includes:

[0010] Data acquisition module: used to acquire the spatio-temporal coordinates (x, y, z, t) of the position to be predicted and transmit them to the state prediction module, where x represents the coordinate of the position to be predicted in the x-axis direction, y represents the coordinate of the position to be predicted in the y-axis direction, z represents the coordinate of the position to be predicted in the z-axis direction, and t represents the time period to be predicted;

[0011] State prediction module: used to expand the three-dimensional unsteady heat conduction equation;

[0012] Used to construct a greenhouse temperature field prediction model and the total loss;

[0013] Used to process the spatio-temporal coordinates of the position to be predicted and output the temperature field T(x′, y′, z′, t′) of the position to be predicted, where x′ represents the coordinate of the temperature field in the x-axis direction, y′ represents the coordinate of the temperature field in the y-axis direction, z′ represents the coordinate of the temperature field in the z-axis direction, and t′ represents the time period when the temperature field is predicted;

[0014] Multi-objective optimization module: used to construct a multi-objective optimization model, and use model predictive control MPC to minimize the power of ventilation equipment and heating equipment within the temperature range [T min , T max that meets the growth of crops, where T min represents the lowest temperature that meets the growth of crops, and T max represents the highest temperature that meets the growth of crops;

[0015] Execution module: used to adjust the operating parameters of the ventilation equipment and heating equipment according to the minimized power of the ventilation equipment and heating equipment, and complete the regulation of the greenhouse temperature field.

[0016] Furthermore, expand the three-dimensional unsteady heat conduction equation: introduce the crop transpiration and dynamic light terms:

[0017] Among them, represents the rate of change of temperature with time, ρ represents the air density, C p represents the specific heat capacity, represents the partial derivative operation symbol, S0 represents the solar constant, f light(t′) represents the variation simulating the daily cycle, g shade () represents the spatial occlusion function of the sunshade curtain, b represents the b-th temperature sample in the crop growth area, β represents the crop transpiration coefficient, ReLU() represents the non-linear activation function, T base represents the base temperature threshold.

[0018] Furthermore, the greenhouse temperature field prediction model adopts a fully connected neural network, including an input layer, a hidden layer and an output layer;

[0019] The input layer includes: the spatio-temporal coordinates (x, y, z, t) of the position to be predicted;

[0020] The hidden layer: The hidden layer has a total of 5 layers and adopts the Swish activation function;

[0021] The output layer adopts a linear activation function;

[0022] The output layer outputs the predicted value: T = f NN (x′, y′, z′, t′; θ), where f NN () represents the neural network function, which is used to predict the temperature field T according to the input spatio-temporal coordinates and network parameters θ, θ ∈ R d , R represents the set of real numbers, and d represents the number of network parameters.

[0023] Furthermore, the total loss of the greenhouse temperature field prediction model where λ i represents the dynamic weight, i = {1, 2, 3}, represents the physical equation residual loss term, represents the sensor loss term, represents the crop constraint loss term.

[0024] Furthermore, the physical equation residual loss term S(x′, y′, z′, t′) = S0·f light (t′)·g shade (x′, y′, z′), S(x′, y′, z′, t′) is the light radiation term, which represents the heat effect caused by light at the spatio-temporal coordinates (x′, y′, z′, t′), Q crop (T) = -β·ReLU(T - T base ), Q crop (T) is the crop transpiration term, which represents the influence of crop transpiration on the temperature in the greenhouse within the temperature field T, D represents the total number of physical equation residual sampling point data sets in the computational domain Ω, d = {1, 2,..., D};

[0025] The sensor loss term where, Denote the actual temperature value measured by the j-th sensor at the position (x j , y j , z j ) and time t j . T(x j ′, y j ′, z j ′, t j ′) represents the predicted temperature field value of the j-th sensor. M represents the total number of sensors for which the prediction error is to be calculated, and j = {1, 2,..., M};

[0026] Crop constraint loss term T b represents the temperature of the b-th temperature sample, and B represents the total number of temperature samples in the crop growth area, where b = {1, 2,..., B}.

[0027] Furthermore, the update rule for the dynamic weight λ i is as follows: n represents the number of iterations of the dynamic weight, μ represents the smoothing coefficient, represents the gradient of the loss term with respect to the network parameter θ, represents the loss term corresponding to λ i , represents the gradient of the total loss with respect to the network parameter θ.

[0028] Furthermore, the multi-objective optimization model is specifically: N t represents the total number of discrete time steps of the multi-objective optimization model, and n t =

[0029] {1, 2,..., N t}, represents the ventilation equipment power at the n t -th time step, represents the heating equipment power at the n t -th time step, and s.t. represents the constraint condition.

[0030] The beneficial effects of the method of the present invention are as follows:

[0031] (1) By integrating the physical laws of thermodynamics, fluid mechanics, and crop transpiration, the present invention establishes PDE constraints including the Navier-Stokes equation and the convection-diffusion equation, accurately describing the multi-field interaction in three-dimensional space.

[0032] (2) Design an encoder-decoder network based on the attention mechanism to enhance the feature extraction of locally sensitive areas (such as animal activity areas); combine real-time sensor data to drive the digital twin system, and generate an optimal energy consumption control strategy based on the prediction results of the greenhouse temperature field prediction model to achieve ventilation-heating linkage regulation, solve the problem of three-dimensional real-time prediction and regulation of the temperature field in complex environments, and is of great significance for improving the comfort of the breeding environment, reducing energy consumption, and promoting agricultural intelligence.

[0033] (3) By integrating the three-dimensional heat conduction equation and sparse sensor data, the present invention constructs a deep learning model with physical constraints, which can reduce the number of sensors while ensuring that the mean absolute error (MAE) of the test set of the model is 0.6°C. This system can be extended to agricultural scenarios such as greenhouse cultivation and warehouse environment management, promoting the wide application of physical information machine learning technology in the optimization of agricultural complex systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the system described in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. 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.

[0036] This embodiment provides a real-time prediction and regulation system for the greenhouse temperature field based on a physics-informed neural network. The schematic diagram of the system is as Figure 1 shown, and the system includes: a data acquisition module, a state prediction module, a multi-objective optimization module, and an execution module.

[0037] Data acquisition module: used to obtain the spatio-temporal coordinates (x, y, z, t) of the position to be predicted and transmit them to the state prediction module, where x represents the coordinate of the position to be predicted in the x-axis direction, y represents the coordinate of the position to be predicted in the y-axis direction, z represents the coordinate of the position to be predicted in the z-axis direction, and t represents the time period to be predicted.

[0038] State prediction module: used to expand the three-dimensional unsteady heat conduction equation;

[0039] used to construct a greenhouse temperature field prediction model and a total loss function;

[0040] It is used to process the spatio-temporal coordinates of the position to be predicted and output the temperature field T(x′, y′, z′, t′) of the position to be predicted, where x′ represents the coordinate in the x-axis direction of the temperature field, y′ represents the coordinate in the y-axis direction of the temperature field, z′ represents the coordinate in the z-axis direction of the temperature field, and t′ represents the time period for which the temperature field is predicted.

[0041] Multi-objective optimization module: It is used to build a multi-objective optimization model and adopt model predictive control MPC to minimize the power of the ventilation equipment and heating equipment within the crop growth temperature range [T min , T max , where T min represents the lowest temperature for crop growth, and T max represents the highest temperature for crop growth.

[0042] Execution module: It is used to adjust the operating parameters of the ventilation equipment and heating equipment according to the minimized power of the ventilation equipment and heating equipment to complete the regulation of the greenhouse temperature field.

[0043] The three-dimensional unsteady heat conduction equation (including convection and radiation terms) is the basis for describing the dynamic changes of the temperature field, which can comprehensively depict the heat transfer mechanism (conduction, radiation, and transpiration) in the greenhouse, cover the multi-physical field coupling effect, and introduce crop transpiration and dynamic light terms into the three-dimensional unsteady heat conduction equation: Among them, represents the rate of change of temperature with time, ρ represents air density, C p represents specific heat capacity, represents the partial derivative operation symbol, S0 represents the solar constant, f light (t′) represents the change of the simulated daily cycle, g shade () represents the spatial occlusion function of the shading curtain. The spatial occlusion range of the shading curtain is from 0% to 100%. 0% means no occlusion, 100% means full occlusion, b represents the bth temperature sample in the crop growth area, β represents the crop transpiration coefficient, ReLU() represents the non-linear activation function, and T base represents the base temperature threshold (usually 25 degrees Celsius). When T > T base , the transpiration rate is linearly related to (T - T base ); ReLU(r) = max(0, r) is the linear rectifier function, where r represents the input of ReLU().

[0044] In this embodiment, S0 = 1361 W / m 2 , f light (t′) = sin(πt / 12), and β = 120 W / (m 3 ·℃).

[0045] The introduction of crop transpiration and dynamic light terms enables the system of the present invention to respond to the external environment in real time (such as sunshade curtain adjustment and increased crop transpiration).

[0046] Construct a greenhouse temperature field prediction model: The greenhouse temperature field prediction model uses a fully connected neural network (MLP), including an input layer, a hidden layer, and an output layer;

[0047] The input layer includes: the spatio-temporal coordinates (x, y, z, t) of the position to be predicted;

[0048] Hidden layer: The hidden layer has a total of 5 layers and uses the Swish activation function;

[0049] The output layer uses a linear activation function;

[0050] The output layer outputs the predicted value: T = f NN (x′, y′, z′, t′; θ), where f NN () represents the neural network function, which is used to predict the temperature field T according to the input spatio-temporal coordinates and network parameters θ (weights and biases), θ ∈ R d , R represents the set of real numbers, and d represents the number of network parameters.

[0051] The network expression of the predicted value is T(x′, y′, z′, t′) = MLP(x′, y′, z′, t′; θ).

[0052] Train the greenhouse temperature field prediction model:

[0053] Randomly sample within the computational domain Ω to obtain the spatio-temporal coordinates (x, y, z, t) of the position to be predicted, and divide them into a training set, a validation set, and a test set.

[0054] Obtain sensor data from the measured data

[0055] Use the Adam optimizer with an initial learning rate of 10 -3 , and decay by 10% every 1000 steps.

[0056] Training steps:

[0057] (1) Pre-training stage: Only use the physical equation residual loss term Initialize the network parameters and sample 10 5 random points to calculate the loss.

[0058] (2) Joint training stage: Add the sensor loss term and the crop constraint loss term to jointly calculate the loss and dynamically adjust the dynamic weight λ i .

[0059] (3) Convergence judgment: When the total loss The change rate is less than 10 -5 Stop training when it reaches this point.

[0060] The core idea of the Physics-Informed Neural Network (PINN) is to combine physical laws (heat conduction equation) with observational data (sensor data) through a loss function. The total loss function consists of three parts: the physical equation residual loss term, the sensor loss term, and the crop constraint loss term.

[0061] The total loss of the greenhouse temperature field prediction model where λ i represents the dynamic weight, i = {1, 2, 3}, represents the physical equation residual loss term, represents the sensor loss term, represents the crop constraint loss term.

[0062] Physical equation residual loss term S(x′, y′, z′, t′) = S0·f light (t′)·g shade (x′, y′, z′), S(x′, y′, z′, t′) is the light radiation term, representing the heat influence caused by light at the spatio-temporal coordinates (x′, y′, z′, t′), Q crop (T) = -β·ReLU(T - T base ), Q crop (T) is the crop transpiration term, representing the influence of crop transpiration on the greenhouse temperature within the temperature field T, D represents the total number of physical equation residual sampling point datasets within the computational domain Ω, d = {1, 2, …, D}. Physical equation residual loss term Forces the network output to satisfy the heat conduction equation, ensuring that the predicted temperature field conforms to the laws of thermodynamics.

[0063] Time derivative and spatial gradient are calculated through automatic differentiation:

[0064]

[0065] Second-order derivative is calculated through the chain rule:

[0066]

[0067] Sensor loss term where, represents the j-th sensor at the position (x j , y j , z j ) and time tj The actual temperature value measured internally, T(x j ′, y j ′, z j ′, t j ′) represents the predicted value of the temperature field of the j-th sensor, M represents the total number of sensors for which the prediction error is to be calculated, j = {1, 2, …, M}; the sensor loss term can minimize the error between the predicted value and the sparse sensor data.

[0068] Crop constraint loss term T b represents the temperature of the b-th temperature sample, B represents the total number of temperature samples in the crop growth area, b = {1, 2, …, B}.

[0069] Crop constraint loss term can penalize the predicted values that exceed the suitable temperature range of the crop.

[0070] To balance the contributions of various losses, a dynamic weight adjustment mechanism is adopted. The weights are dynamically adjusted according to the gradient magnitudes of the loss terms, giving priority to optimizing the terms that are difficult to converge and avoiding a certain loss term from dominating the training (such as overfitting the sensor data and ignoring the physical laws). The dynamic weight adjustment mechanism can improve the convergence speed and stability of the model. Experiments show that the training time in this embodiment is reduced by about 30%.

[0071] The initial weights are λ1 = 0.5, λ2 = 0.3, λ3 = 0.2, and the weights are updated every 1000 steps according to the gradient magnitudes of the respective losses: n represents the number of dynamic weight iterations, μ represents the smoothing coefficient, represents the loss term of the gradient of the network parameter θ, represents λ i corresponding loss term, represents the total loss of the gradient of the network parameter θ.

[0072] Formula solving process:

[0073] Taking the physical equation residual term as an example, its solving process is described in detail:

[0074] (1) Forward propagation:

[0075] Input the spatio-temporal coordinates (x, y, z, t), and calculate the output T(x′, y′, z′, t′) through the greenhouse temperature field prediction model.

[0076] Calculate the time derivative and the spatial gradient

[0077] (2) Residual calculation:

[0078] Calculate the second derivative

[0079] Calculate the light radiation term S(x′, y′, z′, t′) and the crop transpiration term Q crop (T).

[0080] Calculate the physical equation residual:

[0081]

[0082] (3) Loss calculation:

[0083] Calculate the mean square error of the physical equation residual term:

[0084] where W represents the total number of terms of the physical equation residual term, w = {1, 2, …, W}, Residual w represents the w-th physical equation residual term.

[0085] (4) Backpropagation:

[0086] Calculate the total loss The gradient of the network parameter θ Update the network parameters:

[0087] where η represents the learning rate and q represents the number of times the network parameters are iteratively updated.

[0088] Multi-objective optimization model: s.t. T min ≤ T(x′, y′, z′, t′) ≤ T max , N t represents the total number of discrete time steps of the multi-objective optimization model, n t = {1, 2, …, N t} represents the ventilation equipment power at the n t -th time step, represents the heating equipment power at the n t -th time step, s.t. represents the constraint condition.

[0089] Adopt model predictive control MPC to optimize the power of the heating and ventilation equipment within the crop growth temperature range [T min , T max to obtain the minimum power of the heating and ventilation equipment.

[0090] Rolling optimize the power of the ventilation and heating equipment every 10 minutes.

[0091] The present invention constructs a deep learning model with physical constraints by integrating the three-dimensional heat conduction equation and sparse sensor data. While reducing the number of sensors, it ensures that the mean absolute error (MAE) of the test set of the model is 0.6 °C and meets the crop growth requirements (within ±1 °C). Through model predictive control (MPC), the present invention optimizes the operating parameters of ventilation and heating equipment, and reduces energy consumption by 15-20% on the premise that the temperature field distribution meets the crop growth requirements.

Claims

1. A real-time prediction and regulation system for greenhouse temperature field based on physics-informed neural network, characterized in that, The system includes: A data acquisition module: used to acquire the spatio-temporal coordinates (x, y, z, t) of the position to be predicted and transmit them to the state prediction module, where x represents the coordinate of the position to be predicted in the x-axis direction, y represents the coordinate of the position to be predicted in the y-axis direction, z represents the coordinate of the position to be predicted in the z-axis direction, and t represents the time period to be predicted; A state prediction module: used to expand the three-dimensional unsteady heat conduction equation; Used to construct a greenhouse temperature field prediction model and the total loss; Used to process the spatio-temporal coordinates of the position to be predicted and output the temperature field T(x′, y′, z′, t′) of the position to be predicted, where x′ represents the coordinate of the temperature field in the x-axis direction, y′ represents the coordinate of the temperature field in the y-axis direction, z′ represents the coordinate of the temperature field in the z-axis direction, and t′ represents the time period when the temperature field is predicted; Multi-objective optimization module: used to build a multi-objective optimization model, and adopt model predictive control (MPC) to minimize the power of ventilation equipment and heating equipment within the crop growth temperature range [T min , T max , where T min represents the minimum temperature for crop growth, and T max represents the maximum temperature for crop growth; An execution module: used to adjust the operating parameters of the ventilation equipment and the heating equipment according to the minimized power of the ventilation equipment and the heating equipment, and complete the regulation of the greenhouse temperature field.

2. The real-time prediction and regulation system for greenhouse temperature field based on the physics-informed neural network according to claim 1, wherein Extended three-dimensional unsteady heat conduction equation: Introduce crop transpiration and dynamic light terms: Among them, represents the change rate of temperature with time, ρ represents air density, C p represents specific heat capacity, represents the partial derivative operation symbol, S0 represents the solar constant, f light (t′) represents the change of the simulated daily cycle, g shade () represents the spatial occlusion function of the sunshade curtain, b represents the b-th temperature sample in the crop growth area, β represents the crop transpiration coefficient, ReLU() represents the non-linear activation function, T base represents the base temperature threshold.

3. The real-time prediction and control system of the greenhouse temperature field based on the physics-informed neural network according to claim 2, characterized in that, The greenhouse temperature field prediction model adopts a fully connected neural network, including an input layer, a hidden layer, and an output layer; The input layer includes: the spatio-temporal coordinates (x, y, z, t) of the position to be predicted; The hidden layer: The hidden layer has a total of 5 layers and uses the Swish activation function; The output layer uses a linear activation function; The output layer outputs the predicted value: T = f NN (x′, y′, z′, t′; θ), where f NN () represents the neural network function, which is used to predict the temperature field T according to the input spatio-temporal coordinates and the network parameter θ, θ ∈ R d , R represents the set of real numbers, and d represents the number of network parameters.

4. The real-time prediction and regulation system of the greenhouse temperature field based on the physics-informed neural network according to claim 3, characterized in that, Total loss of the greenhouse temperature field prediction model where λ i represents the dynamic weight, i = {1, 2, 3}, represents the physical equation residual loss term, represents the sensor loss term, represents the crop constraint loss term.

5. The real-time prediction and regulation system of the greenhouse temperature field based on the physics-informed neural network according to claim 4, characterized in that Residual loss term of physical equation S(x′, y′, z′, t′) = S0·f light (t′)·g shade (x′, y′, z′), S(x′, y′, z′, t′) is the light radiation term, representing the heat effect caused by light at the space-time coordinates (x′, y′, z′, t′), Q crop (T) = -β·ReLU(T - T base ), Q crop (T) is the crop transpiration term, representing the influence of the crop on the temperature in the greenhouse through transpiration in the temperature field T. D represents the total number of physical equation residual sampling point data sets in the computational domain Ω, d = {1, 2, …, D}; Sensor loss term Among them, represents the actual temperature value measured by the j-th sensor at the position (x j , y j , z j ) and time t j , and T(x′ j , y′ j , z′ j , t′ j ) represents the predicted value of the temperature field of the j-th sensor. M represents the total number of sensors for which the prediction error is to be calculated, and j = {1, 2, …, M}; Crop constraint loss term T b represents the temperature of the b-th temperature sample, B represents the total number of temperature samples in the crop growth area, and b = {1, 2, …, B}.

6. The real-time prediction and regulation system of the greenhouse temperature field based on the physics-informed neural network according to claim 4, wherein Dynamic weight λ i Update rule: n represents the number of iterations of the dynamic weight, μ represents the smoothing coefficient, represents the loss term the gradient of the network parameter θ, represents λ i the corresponding loss term, represents the total loss the gradient of the network parameter θ.

7. The real-time prediction and regulation system of the greenhouse temperature field based on the physics-informed neural network according to claim 1, characterized in that The multi-objective optimization model is specifically as follows: s.t.T min ≤T(x′,y′,z′,t′)≤T max , N t represents the total number of discrete time steps of the multi-objective optimization model, n t ={1, 2, …, N t}}, represents the power of the ventilation equipment at the nth t time step, represents the power of the heating equipment at the nth t time step, s.t. represents the constraint condition.

Images