A greenhouse environment yield and energy consumption combined simulation method based on a hybrid model

Abstract

The application discloses a greenhouse environment yield and energy consumption combined simulation method based on a hybrid model, and comprises the following steps: acquiring multi-source time sequence input data of a greenhouse system; calculating a physical item derivative of the greenhouse system by using a physical model; dynamically generating a weight and calculating a residual item derivative by using a neural network; fusing the physical item derivative and the residual item derivative to obtain a total state derivative; inputting the total state derivative into a common differential equation solver to perform continuous time integration, and outputting indoor microclimate, crop yield and equipment energy consumption simulation results. The application takes into account the explainability of the physical model and the high-precision fitting capability of deep learning, and realizes synchronous high-fidelity combined simulation of the greenhouse environment, yield and energy consumption.

Description

Technical Field

[0001] This invention relates to the intersection of agricultural information technology and artificial intelligence, specifically to a method for joint simulation of greenhouse environment yield and energy consumption based on a hybrid model. Background Technology

[0002] Currently, in the field of greenhouse system simulation, mechanistic-driven physical models or data-driven deep learning models are mainly used. Physical models are built based on energy transfer and mass balance equations and have strong interpretability. However, their prediction accuracy is limited under complex actual operating conditions due to difficulties in parameter measurement and insufficient ability to describe high-frequency nonlinear disturbances. While deep learning models can directly learn nonlinear mapping relationships from historical data, avoiding complex equation derivations, they are essentially black-box models lacking mechanistic constraints. Their performance heavily depends on sample quality, and they are prone to outputting distorted predictions that violate physical laws under operating conditions outside the training set. Therefore, existing single modeling methods cannot simultaneously achieve both physical interpretability and high nonlinear accuracy in greenhouse system simulation, failing to meet the urgent needs of modern intelligent greenhouse management and energy conservation for high-fidelity simulation. Summary of the Invention

[0003] The purpose of this invention is to provide a method for co-simulating greenhouse environment, yield, and energy consumption based on a hybrid model. This invention balances the interpretability of the physical model with the high-precision fitting capability of deep learning, achieving simultaneous high-fidelity co-simulation of greenhouse environment, yield, and energy consumption.

[0004] The technical solution of this invention: a method for joint simulation of greenhouse environment output and energy consumption based on a hybrid model, comprising the following steps: Step S1: Obtain multi-source time-series input data of the greenhouse system, including time-varying external weather data and equipment control action data; Step S2: Calculate the physical derivatives of the greenhouse system based on the multi-source time-series input data using the physical model; Step S3: Utilize a neural network to dynamically generate network weights based on the time-varying external weather data, and calculate the derivative of the residual term based on the prior prediction output of the multi-source time-series input data and the physical model. Step S4: Combine the physical term derivative and the residual term derivative to obtain the total state derivative; Step S5: Input the total state derivative into the ordinary differential equation solver to perform continuous-time integration, and output the simulation results of the greenhouse system. The simulation results include indoor microclimate data, crop yield data, and equipment energy consumption data.

[0005] The above method employs a supernetwork architecture that dynamically generates weights based on external time-varying conditions in the neural network. The supernetwork architecture replaces static network weights with low-dimensional matrix products dynamically generated from the time-varying external weather data, so as to adaptively modulate the coupling relationship between the state variables of the greenhouse system according to environmental disturbances.

[0006] The aforementioned method, in which the hybrid model consisting of the physical model and the neural network is obtained through progressive two-stage training paradigm; In the first phase, only physical models are used to correct uncertain parameters in order to anchor a set of parameters that conforms to the current greenhouse structure and crop characteristics; In the second stage, physical models and neural networks are jointly trained.

[0007] The aforementioned method, wherein the joint training process includes: The hybrid model is forward propagated based on a differentiable integrator, and numerical integration is performed in the time domain to solve for the continuous predicted trajectory. By utilizing a dynamic weighting mechanism based on the Softmax function, the multi-objective loss between the continuous predicted trajectory and the actual observation data is adaptively balanced; A multi-dataset polling update strategy is introduced to maintain the continuity of the time series, and a cosine annealing learning rate decay strategy is combined for gradient backpropagation and parameter optimization.

[0008] The aforementioned method uses a physical model consisting of a climate sub-model, a crop sub-model, and an energy consumption sub-model. The physical model is formalized as a set of first-order continuous-time ordinary differential equations, driven by outdoor meteorological parameters, control action parameters, greenhouse design parameters, and crop physiological parameters, to describe the dynamic changes of the greenhouse system.

[0009] In the aforementioned method, the climate sub-model explicitly introduces the evaporative cooling pad temperature as an input variable to calculate the dynamic cooling efficiency of the fan-evaporative cooling pad system; The formula for calculating sensible heat flux in the climate sub-model is as follows: ; In the formula, For sensible heat flux, The ventilation flux generated by the fan-evaporative cooling pad system. air density, The specific heat capacity of air at constant pressure. Outdoor dry-bulb temperature, The latent heat of vaporization of water, For dynamic cooling efficiency, This refers to the moisture content of the evaporative cooling pad. The humidity content of outdoor air; The formula for calculating the steam flux under dynamic changes in indoor humidity is: ; In the formula, For steam flux, For the efficiency of the wet curtain; Ventilation flux of the fan-evaporative cooling pad system The calculation formula is: ; In the formula, The opening degree of the fan-evaporative cooling pad system; The ventilation capacity of the evaporative cooling pad; The greenhouse floor area; According to the equation for moist air in an adiabatic saturated process Represented as: ; In the formula, Represents the outdoor dry-bulb temperature. It serves as the thermodynamic wet-bulb temperature. The saturated moisture content of the evaporative cooling pad is given by the following formula: ; In the formula, The saturated vapor pressure at the wet curtain temperature. Absolute atmospheric pressure; The dynamic cooling efficiency The calculation formula is: ; In the formula, This refers to the temperature of the evaporative cooling pad. It is the thermodynamic wet-bulb temperature.

[0010] The aforementioned method uses a crop sub-model based on a source-sink mechanism to simulate the process of carbohydrates from production and storage to conversion into crop organ dry weight. The formulas for calculating carbohydrate allocation fluxes in fruits, leaves, and stems in the crop sub-model are as follows: ; ; ; In the formula, As a carbohydrate deficiency inhibitor, As a factor that inhibits instantaneous temperature discomfort, It is an inhibitory factor for discomfort caused by 24-hour average temperature. It is an inhibitory factor that is unsuitable for crop development stages. The effect of temperature on carbohydrate flux in fruit. , , These are the potential growth rate coefficients for fruits, leaves, and stems, respectively.

[0011] The aforementioned method, in the crop sub-model, uses the following calculation formulas for maintenance respiration and growth respiration: The formula for maintaining respiration is: ; In the formula, The weight of carbohydrates in plant organs. The maintenance respiration coefficient of plant organs. The effect of temperature on maintaining respiration value, The average canopy temperature over 24 hours. To maintain the respiratory regression coefficient, Relative growth rate; This represents the relative growth rates of three different types of organic matter. relative growth rate The calculation formula is: ; In the formula, For the maximum relative growth rate, This is the cumulative amount of daily radiation. It is the Michaelis constant; The formula for calculating growth and respiration is: ; In the formula, This represents the growth and respiration coefficient of plant organs; For the first The carbohydrate buffer capacity of plant organs.

[0012] The aforementioned method, wherein the energy consumption sub-model includes heating energy consumption and cooling energy consumption, is used to quantify the electrical energy consumed by the greenhouse system in maintaining the target microclimate; The formula for calculating heating energy consumption is: ; In the formula, For the control ratio of the boiler, For boiler power, The greenhouse floor area For boiler efficiency; The formula for calculating cooling energy consumption is: ; In the formula, For cooling load per unit area, This refers to the rated power of the fan-evaporative cooling pad system.

[0013] The aforementioned method, after acquiring the multi-source time-series input data of the greenhouse system, further includes: Data outliers are removed based on the boundary conditions of each parameter. The SAITS algorithm is used to impute missing data in order to obtain continuous and complete time series data.

[0014] Compared with existing technologies, this invention constructs a hybrid architecture coupling a physical model and a neural network. The physical model is used to determine the main changing trends of the greenhouse system, ensuring that the simulation process follows fundamental physical laws such as energy and matter conservation, thus providing a solid foundation for the model's interpretability. Simultaneously, by using the neural network to dynamically generate weights based on time-varying external conditions, it can adaptively capture and compensate for unknown dynamic deviations that are difficult to accurately describe under complex unsteady conditions, achieving a deep fusion of physical mechanisms and data-driven features. Furthermore, by inputting the fused total state derivative into the ordinary differential equation solver for continuous-time integration, the continuity and smoothness of the simulation trajectory in the time domain are ensured, avoiding the cumulative errors caused by discrete time steps. In addition, the progressive two-stage training paradigm and multi-dataset polling update strategy ensure the reliability of the physical meaning of the physical parameters and resolve the contradiction between the time series continuity requirement in ODE model training and the traditional random sampling strategy. This significantly improves the prediction accuracy and generalization robustness of the hybrid model under variable environments, ultimately achieving synchronous high-fidelity co-simulation of greenhouse microclimate, crop yield, and equipment energy consumption. Attached Figure Description

[0015] Figure 1 This is a flowchart of the simulation method of the present invention; Figure 2 This is a schematic diagram of the heat and mass exchange architecture of the greenhouse physical model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture and training process of the hybrid model of the present invention; Figure 4 This is a schematic diagram of the hybrid model simulation process of the present invention. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0017] Example: This example provides a co-simulation method for greenhouse environment output and energy consumption based on a hybrid model. During the inference and execution phase, this method achieves high-fidelity continuous-time simulation of the dynamic changes in the greenhouse system through deep integration of physical mechanisms and data-driven features. Specifically, as follows... Figure 1 As shown, the simulation method includes the following steps: Step S1: Obtain multi-source time-series input data of the greenhouse system, including time-varying external weather data and equipment control action data.

[0018] Specifically, time-varying external weather data refers to external meteorological stimuli that drive changes in the greenhouse environment, such as parameters like solar radiation, atmospheric temperature, relative humidity, carbon dioxide concentration, wind speed, sky temperature, and soil temperature. Equipment control action data refers to control commands applied to the greenhouse system, such as boiler heating ratio, fan-evaporative cooling system opening degree, top / side window ventilation opening degree, and shading curtain deployment ratio. In this embodiment, the aforementioned multi-source time-series input data must be strictly aligned on the time axis to serve as the synchronization driving source for the subsequent physical model and neural network. It should be understood that although this embodiment lists specific meteorological and control parameter types, in other implementations, specific input variables can be added or reduced according to the specific greenhouse configuration or simulation accuracy requirements, as long as they can characterize the external stimuli affecting the evolution of the greenhouse state.

[0019] After acquiring multi-source time-series input data from the greenhouse system, a data preprocessing step is included to ensure the reliability of subsequent physical model solving and neural network training. Specifically, outliers are first eliminated based on the boundary conditions of each parameter. These "boundary conditions" are not simple statistical thresholds (such as the 3σ principle), but rather reasonable constraints based on basic greenhouse physics. For example, indoor relative humidity cannot exceed 100% or fall below 0%, CO2 concentration under normal ventilation conditions should not be lower than the outdoor background value (approximately 400 ppm) or higher than the upper limit of supplementary gas (such as 2000 ppm), and the rate of temperature change should not exhibit abrupt changes of tens of degrees per second under thermal inertia. Through this anomaly detection based on physical boundaries, false data points caused by sensor drift, communication packet loss, or electromagnetic interference can be effectively identified and eliminated, preventing these noises that violate physical laws from contaminating the model parameter calibration process. It should be understood that for data points determined to be outliers, this embodiment marks them as missing rather than directly deleting the entire record, in order to preserve the integrity of the timeline.

[0020] For data marked as missing, as well as truly missing data caused by equipment maintenance during the original data acquisition process, the SAITS algorithm is used to imputate the missing data to obtain continuous and complete time series data. SAITS (Self-supervised Attention-based Imputation for Time Series) is a time series imputation algorithm based on a self-supervised attention mechanism. Compared to traditional linear interpolation, spline interpolation, or mean imputation methods, the core advantage of the SAITS algorithm lies in its ability to capture long-range dependencies and dynamic correlations between multidimensional variables within the time series. In a greenhouse scenario, variables such as temperature, humidity, and radiation are tightly coupled physically and exhibit significant diurnal and seasonal patterns. When temperature data is missing at a certain moment, SAITS can not only use temperature information from adjacent moments for a smooth transition but also simultaneously reference radiation intensity, humidity changes during the same period, and temperature patterns under similar historical conditions to infer the imputation value that best matches the current physical state. This high-quality imputation is crucial to this invention because ordinary differential equation (ODE) solvers require state variables to be continuous and smooth in the time domain. If simple linear interpolation is used, it will introduce sharp points of derivative discontinuity at both ends of the missing segment, causing numerical oscillations or accumulation of integral errors in the ODE solver at that point. However, the interpolation curves generated by SAITS are highly consistent with the actual physical process in terms of trend, maintaining the dynamic continuity of time series data to the greatest extent, and laying a solid data foundation for high-fidelity simulation of hybrid models.

[0021] Step S2: Calculate the physical derivative of the greenhouse system based on the multi-source time-series input data using the physical model.

[0022] Specifically, the physical model is a mechanistic model built upon the energy transfer processes (such as conduction, convection, and radiation) and the law of conservation of mass (such as water vapor and carbon dioxide flux) within the greenhouse. During simulation, the physical model receives current multi-source time-series input data and calculates the instantaneous rate of change of various greenhouse state variables (such as indoor air temperature, humidity, and crop biomass) at the current moment through analytical equations; this is the physical derivative. This physical derivative reflects the deterministic evolution trend of the greenhouse system under the constraints of ideal physical laws, providing a baseline trajectory that conforms to the basic principles of thermodynamics and biology for the entire simulation process, ensuring that the simulation results do not deviate from basic physical common sense.

[0023] Step S3: Utilize the neural network to dynamically generate network weights based on the time-varying external weather data, and calculate the derivative of the residual term based on the multi-source time-series input data and the prior prediction output of the physical model.

[0024] Specifically, due to the numerous complex factors in actual greenhouse systems that are difficult to accurately describe using analytical equations, such as the nonlinear degradation of equipment performance over time, turbulence effects under unsteady ventilation, or transient responses of crop stomatal conductance, relying solely on physical models often leads to cumulative biases. This embodiment introduces a neural network to specifically fit these unknown dynamic characteristics. During inference, the neural network does not use fixed static weights but instead uses time-varying external weather data as input, dynamically adjusting its internal network connection weights in real time to adaptively perceive the current environmental conditions. Simultaneously, the neural network also receives multi-source time-series input data and prior predictions from the physical model as features, calculating the derivative of the residual term used to correct physical biases. This dynamic weight generation mechanism allows the residual compensation capability to be flexibly modulated according to changes in the external environment, effectively improving the model's adaptability under unsteady conditions. It should be noted that this step focuses on describing the input-output behavior and dynamic adjustment function of the neural network during simulation. The specific hypernetwork architecture and weight generation mechanism within the neural network will be elaborated in detail below.

[0025] Step S4: Combine the physical term derivative and the residual term derivative to obtain the total state derivative.

[0026] Specifically, the physical derivative obtained in step S2 is superimposed and fused with the residual derivative obtained in step S3. The physical derivative forms the main framework of the system dynamics, ensuring the physical interpretability of the simulation; the residual derivative serves as a high-precision correction term, compensating for the insufficient description of the physical model in local nonlinear regions. The resulting total state derivative retains the robustness of the mechanistic model while possessing the ability of a data-driven model to capture complex dynamic characteristics, forming a complete mathematical description of the true evolution of the greenhouse system.

[0027] Step S5: Input the total state derivative into the ordinary differential equation solver to perform continuous-time integration, and output the simulation results of the greenhouse system. The simulation results include indoor microclimate data, crop yield data, and equipment energy consumption data, such as... Figure 4 As shown.

[0028] Specifically, the Ordinary Differential Equation (ODE) solver receives the total state derivative and performs numerical integration within a preset time interval. Unlike traditional discrete-time step prediction models, the ODE solver can solve for the evolution trajectory of state variables in the continuous time domain. This not only avoids the accumulation of truncation errors caused by fixed step sizes but also naturally ensures the smoothness and continuity of the simulation curve, which is more consistent with the physical nature of the heat and mass exchange process in greenhouses. The final simulation results cover three key dimensions: indoor microclimate data (such as air temperature and humidity, CO2 concentration, etc.), crop yield data (if the actual cumulative fresh weight), and equipment energy consumption data (such as the power consumption of wet curtains and fans). Through the above execution process, this invention achieves synchronous high-fidelity co-simulation of greenhouse environment, crop growth, and energy consumption at a unified time scale, providing reliable data support for intelligent management and energy-saving optimization of greenhouses.

[0029] Figure 3 The diagram illustrates the overall modeling process, which is divided into five functional modules: model input, physical model, hypernetwork neural network, differentiable ODE solver, and backpropagation optimization. The input parameters of the model input module are uniformly shared with the physical model and neural network, and are specifically divided into three categories: first, outdoor meteorological parameters (i.e., outdoor weather in the diagram), including solar radiation, atmospheric temperature, atmospheric relative humidity, atmospheric CO2 concentration, wind speed, soil temperature, and total daily solar radiation; second, facility control parameters (i.e., control actions in the diagram), corresponding to control commands for four types of actuators: boiler, wet curtain-fan, ventilation window, and shading curtain; and third, inherent parameters of the greenhouse configuration (i.e., greenhouse design parameters in the diagram), including greenhouse enclosure dimensions, ventilation characteristics, and radiation and heat transfer characteristics of each component (roof, shading curtain, floor, and soil).

[0030] The physical model module consists of a climate sub-model, a crop sub-model, and an energy consumption sub-model coupled together, which correspond one-to-one with the aforementioned greenhouse multi-field coupled topology. The three types of sub-models obtain the mechanism state derivatives by solving the energy conservation, water vapor, and mass conservation equations. The mechanism state derivatives are then fed down into the summation node. At the same time, the physical model initially outputs three types of predicted data: microclimate, yield, and energy consumption (i.e., the predicted data in the figure).

[0031] The input data of the neural network module is first preprocessed by a normalization layer (Norm, i.e., standardization) and then divided into two feature extraction branches. Each branch sequentially passes through a fully connected layer (Linear) and the SiLU activation function to complete feature extraction. The extraction results are decomposed into matrices A and B by low-rank decomposition, and feature fusion is achieved by matrix multiplication. The fused data is then normalized, activated, and fully connected again to output the derivative of the residual term (res, i.e., the residual term). The derivative of the residual term is fed into the summation node and added to the aforementioned mechanism state derivative to obtain the mixed derivative. At the same time, the neural network outputs the network weight parameters to the right optimization unit.

[0032] The differentiable ODE solver module uses the Euler method (i.e., Euler numerical integration algorithm) as the integration method. After receiving the mixed derivative, it performs time-series integration operations, outputs the optimized simulation results of microclimate, yield, and energy consumption, and sends the simulation data to the optimization unit on the right.

[0033] The backpropagation and optimization module matches and verifies the model's predicted data with the measured observations, and uses a batch input strategy of polling the dataset (i.e., polling update in the graph). Each sub-loss is dynamically weighted by Softmax (i.e., dynamic weighted loss) and accumulated to construct a multi-objective loss function. Backpropagation is performed based on the gradient descent algorithm to simultaneously complete the iterative update of two types of parameters: neural network weights and uncertain parameters of the physical model. The updated parameters are then looped back to the neural network and the physical model to achieve full-model iterative optimization.

[0034] Figure 3 The solid line represents the first stage, and the dashed line represents the second stage. In the first stage, only the physical model is used for parameter calibration: the physical model receives the model input, calculates the state derivative based on the mechanistic equations, integrates it through the ODE solver, and outputs the prediction result. Then, the backpropagation optimization module continuously corrects the uncertain parameters. In the second stage, the physical model and the neural network are jointly trained. The calibrated physical model receives the model input, calculates the state derivative of the physical term based on the mechanistic equations, and integrates it through the ODE solver to output the prior prediction. The model input and the prior prediction are simultaneously input into the neural network, outputting the derivative of the residual term. The derivatives of the residual term and the physical term together constitute the mixed derivative, which is integrated through the ODE solver to output the prediction result. Then, the backpropagation optimization module continuously updates the neural network parameters.

[0035] Furthermore, this embodiment provides a detailed description of the internal architecture of the neural network and its weight generation mechanism. In this embodiment, the neural network adopts a hypernetwork architecture that dynamically generates weights based on external time-varying conditions. Specifically, traditional deep learning models typically use static weight matrices that are fixed after being determined during the training phase. This static parameterization method assumes that the dynamic characteristics of the system are stationary or approximately stationary throughout the entire operating cycle. However, a greenhouse system is a typical strongly nonlinear, time-varying dynamic system, and its internal heat and mass exchange processes exhibit significant non-steady-state characteristics due to the drastic influence of outdoor weather conditions. To overcome the limitations of static networks, this invention introduces the concept of a hypernetwork, constructing an adaptive compensation module that can dynamically adjust its own parameters based on real-time environmental perception. In this architecture, the hypernetwork replaces the static network weights with low-dimensional matrix products dynamically generated from time-varying external weather data, so as to adaptively modulate the coupling relationship between the state variables of the greenhouse system according to environmental disturbances.

[0036] Specifically, the aforementioned "low-dimensional matrix multiplication" is essentially a low-rank factorization mechanism. During inference, the weights of the backbone layer of the neural network are no longer constant tensors directly read from memory, but are calculated in real time by a lightweight generator network. This generator network takes current time-varying external weather data (such as outdoor temperature, humidity, radiation intensity, etc.) as input and outputs a set of low-dimensional vectors or matrix factors. These factors are used to reconstruct the high-dimensional weight matrix required by the backbone network at the current moment through outer product or linear combination operations. It should be understood that although this embodiment uses low-rank matrix multiplication as the preferred implementation, in other embodiments, dynamic convolution kernel generation, conditional affine transformation, or other parameterized mechanisms that can achieve "input-dependent weight generation" can also be used, as long as their core logic is to use external environmental information to adjust the signal transmission path inside the network in real time.

[0037] This dynamic weight generation mechanism endows the neural network with extremely high physical flexibility, enabling it to adaptively modulate the coupling relationships between state variables according to environmental disturbances. From a microscopic perspective, the coupling strength between various state variables (such as temperature, humidity, and CO2 concentration) inside the greenhouse is not constant but dynamically drifts with external conditions. For example, on a sunny afternoon in summer with high temperature and humidity, solar radiation is strong, and the fan-evaporative cooling system operates at high frequency. At this time, there is a strong evaporative cooling coupling effect between indoor air temperature and humidity, and the thermal inertia is relatively small. In contrast, at night or on cloudy days in winter, the system mainly relies on heating to maintain the temperature, and the ventilation volume is extremely low. At this time, the coupling between temperature and humidity weakens, while the coupling of state variables related to heat conduction in the building envelope strengthens. If a static weight network is used, the model can only learn the "average" coupling patterns under these various conditions, leading to underfitting or overfitting under extreme or specific conditions. In contrast, the hypernetwork architecture of this embodiment can use external meteorological conditions as a "context index" to dynamically reorganize the connection weights within the network. When high temperature and high humidity conditions are detected, the generator automatically increases the weight gain of feature channels related to latent heat exchange, enhancing the extraction of temperature and humidity coupling features. When low temperature and low light conditions are detected, it automatically suppresses the weights related to convective heat transfer, instead enhancing the response to thermal inertia and radiative heat gain. This triple binding mechanism of "environment, weight, and state" transforms the neural network from a blind black-box fitter into a dynamic compensator capable of self-reconstruction as the environment evolves. This significantly improves the simulation accuracy for complex unsteady conditions while ensuring physical interpretability.

[0038] Furthermore, the dynamic generation strategy using low-dimensional matrix multiplication offers advantages in parameter efficiency. Since the low-dimensional factors output by the generator are much smaller than the actual weights of the backbone network, this significantly reduces the total number of independent parameters the model needs to learn, effectively mitigating the risk of overfitting in small-sample greenhouse data scenarios. Simultaneously, the low-rank constraint itself acts as a regularization mechanism, forcing the model to learn the most sensitive and physically meaningful latent factors to environmental changes, rather than memorizing high-frequency noise in the data. Therefore, this embodiment successfully solves the technical challenge of traditional static neural networks adapting to the time-varying dynamics of greenhouses by combining supernetworks and low-rank decomposition, providing crucial structural support for high-fidelity simulation of hybrid models.

[0039] Furthermore, such as Figure 3 As shown, the hybrid model composed of the physical model and the neural network is obtained through a progressive two-stage training paradigm. This phased training strategy is not a simple engineering choice, but rather determined by the coupling characteristics of the physical mechanisms of the greenhouse system and the data-driven model. Specifically, in the first stage, only the physical model is used to correct uncertain parameters to anchor a parameter set that conforms to the current greenhouse structure and crop characteristics. In this stage, the neural network is in a frozen or inactive state, and the optimizer only optimizes key uncertain parameters in the physical model (such as the convective heat exchange coefficient of the enclosure structure, leakage coefficient, far-infrared emissivity of the roof glass, and potential growth rate of organs). The core purpose of this is to establish a "benchmark base" with clear physical meaning for the hybrid model. If the physical parameters and neural network weights are directly trained end-to-end, due to the powerful nonlinear fitting ability of the neural network, the model can easily "mask" the errors of the physical parameters by adjusting the network weights, resulting in the final physical parameters deviating from the true physical laws, thus losing the interpretability advantage of the hybrid model. The independent calibration in the first stage ensured that the physical model itself could capture the most basic energy and matter conservation trends of the greenhouse system, providing a reliable prior reference for the subsequent intervention of the neural network.

[0040] In the second stage, joint training is performed using both the physical model and the neural network. At this point, the physical parameters are within a reasonable physical range, and the neural network is activated and begins to learn unknown dynamic deviations that the physical model cannot describe. To ensure the stability and effectiveness of the joint training, this embodiment designs a training process specifically adapted to the characteristics of ordinary differential equations (ODEs). Specifically, the joint training process includes the following three core steps: Step S31 involves forward propagating the hybrid model using a differentiable integrator, performing numerical integration in the time domain to solve for the continuous predicted trajectory. Unlike traditional discrete-time step prediction models, the hybrid model of this invention is essentially a neural differential equation system. To simultaneously optimize physical parameters and network weights using gradient descent, the ODE solver used must be differentiable. During forward propagation, the differentiable integrator not only outputs the predicted trajectory of the state variables but also records the computational graph or adjoint state during integration, allowing the error gradient to penetrate back along the time integration path and accurately propagate back to the physical model parameters and neural network weights. This achieves true end-to-end joint optimization of "physics and data," rather than a simple two-stage sequential concatenation.

[0041] Step S32 utilizes a dynamic weighting mechanism based on the Softmax function to adaptively balance the multi-objective loss between the continuous predicted trajectory and the actual observed data. Greenhouse simulation involves multiple output variables such as indoor temperature, relative humidity, CO2 concentration, crop yield, and equipment energy consumption. These variables not only have different physical dimensions but also significantly different numerical ranges and error sensitivities. For example, small deviations in temperature may amplify the crop growth model, while humidity has a relatively high absolute error tolerance. Using fixed weighting coefficients often requires significant manual debugging time and is difficult to adapt to the dynamic changes in the convergence speed of each subtask during training. This embodiment introduces a dynamic weighting mechanism based on the Softmax function, using the instantaneous loss value of each subtask as input and generating dynamic weights through Softmax normalization. When the loss of a certain subtask (such as yield prediction) is significantly higher than other tasks, its corresponding weight automatically increases, forcing the optimizer to prioritize reducing the error of that task; conversely, when a task has converged to a low level, its weight automatically decreases to avoid overfitting. This adaptive balancing mechanism effectively solves the multi-objective optimization problem in multi-physics coupled simulation and improves the overall convergence efficiency of the model.

[0042] Step S33 introduces a multi-dataset polling update strategy to maintain the continuity of the time series, and combines it with a cosine annealing learning rate decay strategy for gradient backpropagation and parameter optimization. This is a key feature that distinguishes this embodiment from conventional deep learning training methods. In traditional neural network training, training samples are usually randomly shuffled to break the correlation between data and prevent overfitting. However, for ODE-based hybrid models, the evolution of state variables has strict temporal causal dependencies. Random shuffling will cause the integration path to break frequently on the time axis, making it impossible for the ODE solver to construct a continuous differential trajectory, thus generating huge gradient noise and even causing numerical divergence. To solve this contradiction, this invention proposes a multi-dataset polling update strategy. Specifically, the original long-term time series data is divided into multiple subsets (slices) that maintain internal temporal continuity according to working conditions or time periods. In each training iteration, the model completely traverses one subset for forward integration and loss calculation, ensuring that the gradient signal comes from a continuous physical evolution process; in the next iteration, it rotates to another subset. This strategy rigorously satisfies the rigid requirement of temporal continuity in ODE solving while introducing necessary sample diversity through polling across different operating condition segments, effectively preventing overfitting of the model to a single continuous segment. Simultaneously, the cosine annealing learning rate decay strategy ensures that the learning rate decreases periodically throughout the training period. This strategy synergizes well with multi-dataset polling: initially, a higher learning rate helps the model quickly adapt to new dynamic characteristics when switching to a new operating condition segment; as training progresses in that segment, the learning rate gradually decreases, helping the model finely search for the optimal solution on the loss surface of the current operating condition. The combination of these two approaches significantly improves the generalization ability and parameter robustness of the hybrid model in complex and variable greenhouse environments.

[0043] It should be understood that although this embodiment uses "multi-dataset polling" as the preferred strategy to maintain temporal continuity, other embodiments may also employ methods such as sliding window sampling or segmented sampling triggered by physical events. As long as the core logic introduces sample diversity while ensuring the continuity of the ODE integration path, these methods fall within the scope of this invention. Similarly, for dynamic weighting mechanisms, in addition to Softmax, adaptive multi-task learning strategies such as uncertainty-based weighting or gradient magnitude normalization (GradNorm) can also be used to achieve similar technical effects.

[0044] Furthermore, in this embodiment, as Figure 3As shown, the physical model consists of a climate sub-model, a crop sub-model, and an energy consumption sub-model. These three sub-models do not operate independently but are tightly coupled through energy flow, material flow, and information flow, forming a holistic framework describing the complex dynamic behavior of the greenhouse system. Specifically, the physical model is formalized as a system of first-order continuous-time ordinary differential equations, driven by outdoor meteorological parameters, control action parameters, greenhouse design parameters, and crop physiological parameters to describe the dynamic changes of the greenhouse system. This formal modeling approach unifies the heat and mass exchange process, crop growth and development process, and equipment energy consumption process within the greenhouse as derivative functions of state variables with respect to time, enabling the model to accurately capture the instantaneous evolution trend of the system at any given moment, rather than merely a static mapping of discrete time points. It should be understood that although this embodiment divides the physical model into three functional sub-modules, in actual implementation, finer-grained divisions (such as further splitting the climate sub-model into air domain, cover layer domain, soil domain, etc.) or coarser-grained integration can be adopted according to the simulation granularity or computational resource requirements. As long as it is essentially a continuous-time dynamic system built based on physical conservation laws, it falls within the protection scope of this invention.

[0045] Among the components of the physical model, the climate sub-model is the core element determining the accuracy of the microenvironment simulation. Addressing the issue in existing technologies where the cooling efficiency of the fan-evaporative cooling pad system is typically simplified to a fixed constant, leading to significant prediction deviations under non-standard operating conditions such as high temperature and low humidity or low temperature and high humidity, this embodiment explicitly introduces the evaporative cooling pad temperature as an input variable to calculate the dynamic cooling efficiency of the fan-evaporative cooling pad system. The physical essence of this improvement lies in the fact that the heat and mass transfer performance of the evaporative cooling pad is not constant but strongly depends on the nonlinear balance between the water film temperature on the evaporative cooling pad surface, the air intake conditions, and the evaporation rate. By adjusting the evaporative cooling pad temperature... As explicit state variables or key inputs, the model can sense and reflect the dynamic drift of cooling efficiency in real time, thereby significantly improving the accuracy of indoor temperature and humidity prediction.

[0046] Specifically, the key physical quantities involved in heat and mass exchange in the fan-evaporative cooling pad system in the climate sub-model are calculated as follows: First, the formula for calculating sensible heat flux is: ; In the formula, For sensible heat flux; This refers to the ventilation flux generated by the fan-evaporative cooling pad system. air density; The specific heat capacity of air at constant pressure; Outdoor dry-bulb temperature; The latent heat of vaporization of water; For dynamic cooling efficiency; This refers to the moisture content of the evaporative cooling pad. This refers to the humidity content of outdoor air.

[0047] Secondly, the formula for calculating the steam flux due to dynamic changes in indoor humidity is: ; In the formula, For steam flux, The formula represents the evaporative cooling efficiency; it reveals that the humidity of the exhaust air is determined by both the theoretical setpoint and the actual cooling efficiency, rather than by simple boundary condition assignment.

[0048] Third, the ventilation flux of the fan-evaporative cooling pad system The calculation formula is: ; In the formula, This represents the opening degree of the fan-evaporative cooling system, with a value ranging from 0 to 1. The ventilation capacity of the evaporative cooling pad; The greenhouse floor area is used to normalize the absolute ventilation volume to the ventilation flux per unit area, facilitating model migration between greenhouses of different sizes.

[0049] Fourth, according to the equation for moist air in an adiabatic saturated process, Represented as: ; In the formula, Represents the outdoor dry-bulb temperature. It serves as the thermodynamic wet-bulb temperature. The saturated moisture content of the evaporative cooling pad is given by the following formula: ; In the formula, Let be the saturated vapor pressure at the evaporative cooling pad temperature, and its value changes exponentially with the evaporative cooling pad temperature. Atm / sq. This parameter defines the maximum amount of moisture air can hold at the current evaporative cooling pad temperature, and is a key thermodynamic benchmark for calculating the evaporation driving force and cooling limit. Calculated according to Tetens' empirical formula: ; The formula is calculated based on the ideal gas law and Dalton's law of partial pressures: ; In the formula, and These are the molar masses of water vapor and dry air, respectively. This is the vapor pressure of outdoor air.

[0050] Finally, and most importantly, the most crucial improvement in this embodiment is the dynamic cooling efficiency. The calculation formula is: ; In the formula, This refers to the temperature of the evaporative cooling pad. The wet-bulb temperature is given by the following formula: ; The dynamic cooling efficiency formula defines the ratio of actual temperature drop to the theoretical maximum temperature drop (i.e., the dry-bulb / wet-bulb temperature difference). In traditional models, this is often assumed to be a fixed empirical value such as 0.7 or 0.8. However, in actual operation, when the evaporative cooling pad water temperature rises due to heating by circulating water, the saturated vapor pressure increases, leading to a weakening of the evaporation driving force, an increase in the actual outlet air temperature, and consequently a significant decrease. Conversely, during dry and cool periods, the evaporative cooling pad efficiency may approach the theoretical extreme value. This embodiment transforms a static parameter into a real-time dynamically changing variable through the aforementioned formula chain. This dynamic characteristic directly corrects the calculation results of the aforementioned sensible heat flux and steam flux, thereby avoiding microclimate simulation distortion caused by efficiency misjudgment under extreme weather conditions.

[0051] It should be noted that although this embodiment provides a specific efficiency calculation formula based on thermodynamic wet-bulb temperature, in other embodiments, equivalent expressions based on Merkel number, Lewis factor, or other heat and mass transfer analogies can also be used to calculate dynamic cooling efficiency. As long as the core logic is to model the cooling efficiency as a function of the thermal state of the wet curtain (such as temperature and enthalpy) rather than a fixed constant, it should be regarded as an equivalent substitution of the inventive concept.

[0052] Furthermore, in this embodiment, the crop sub-model is based on a source-sink mechanism to simulate the process of carbohydrate production, storage, and conversion into crop organ dry weight. Unlike complex plant physiological models that focus on morphological and structural descriptions, the source-sink mechanism used in this embodiment focuses on the macroscopic dynamics of dry matter accumulation and distribution, treating the crop as a dynamic equilibrium system composed of a "source" (photosynthetic production end), a "sink" (organ growth consumption end), and a "buffer pool" (temporary storage end). Specifically, assimilates synthesized through photosynthesis first enter the carbohydrate buffer pool, and then, driven by sink strength, are distributed to organs such as fruits, leaves, and stems. The advantage of this modeling approach is that it avoids the tedious parameterization of microscopic processes such as cell division and elongation, directly linking environmental factors to yield, making it particularly suitable for the simulation needs of greenhouse production management. It should be understood that although this embodiment uses tomato as an example, this source-sink architecture is also applicable to other fruit and vegetable greenhouse crops such as cucumbers and bell peppers, requiring only adjustments to the corresponding distribution coefficients and organ potential growth rates.

[0053] To accurately quantify assimilate allocation under non-ideal conditions, the formulas for calculating carbohydrate allocation fluxes in fruits, leaves, and stems in the crop sub-model are as follows: ; ; ; In the formula, As a carbohydrate deficiency inhibitor, As a factor that inhibits instantaneous temperature discomfort, It is an inhibitory factor for discomfort caused by 24-hour average temperature. It is an inhibitory factor that is unsuitable for crop development stages. The effect of temperature on carbohydrate flux in fruit. , , These are the potential growth rate coefficients for fruits, leaves, and stems, respectively.

[0054] Furthermore, the accuracy of respiration consumption calculation directly determines the accuracy of net biomass accumulation. The calculation formulas for maintenance respiration and growth respiration in the crop sub-model are as follows: The formula for maintaining respiration is: ; In the formula, The weight of carbohydrates in plant organs. The maintenance respiration coefficient of plant organs. The effect of temperature on maintaining respiration The value, usually around 2.0, indicates that the respiration rate doubles for every 10°C increase in temperature. The average canopy temperature over 24 hours. To maintain the respiratory regression coefficient, Relative growth rate; This represents the relative growth rates of three different types of organic matter. The most crucial feature of this formula is the introduction of a dynamic adjustment term. In traditional models, the maintenance respiration coefficient is often treated as a constant related only to temperature and biomass. This leads to the model calculating excessively high respiration consumption in the later stages of crop growth or maturity, despite a significant decrease in tissue metabolic activity, thus underestimating final yield. This invention recognizes that maintenance respiration is essentially providing energy to living cells to maintain vital activities such as ion gradients and protein turnover, and its intensity should be positively correlated with the metabolic activity of the tissue. The relative growth rate... It is the best physiological indicator for characterizing tissue metabolic activity. Through this exponential decay function, when... At higher levels (young tissues), the respiration coefficient is maintained close to its maximum value; when When the respiratory coefficient approaches zero (in aging or mature tissues), it automatically decreases to an extremely low level. This dynamic correction mechanism based on physiological state effectively solves the problem of respiratory estimation bias throughout the reproductive lifespan.

[0055] Let be a rectangular hyperbolic function of the average daily solar radiation integral. Specifically, the relative growth rate of crops increases non-linearly with increasing received radiation, gradually approaching its physiological saturation limit under strong light. Relative growth rate The calculation formula is: ; In the formula, For the maximum relative growth rate, This is the cumulative amount of daily radiation. is the Michaelis constant.

[0056] The formula is described using a right-angled hyperbola. The relationship is with the cumulative amount of solar radiation, rather than a simple linear proportion. This is because the photosynthetic apparatus of crops exhibits light saturation, and under low light intensity... It increases approximately linearly with increasing radiation, but tends to saturate under high light intensity due to limitations in enzyme activity and electron transport rate. The introduction of parameters enables the model to distinguish the differences in growth characteristics between shade-tolerant and light-loving varieties, thereby enhancing the model's universality.

[0057] The formula for calculating growth and respiration is: ; In the formula, This represents the growth and respiration coefficient of plant organs; For the first The carbohydrate buffer capacity of plant organs.

[0058] Growth respiration exhibits a linear relationship with substrate supply, consistent with biochemical reactive scalar principles. Through the synergy of the aforementioned source-sink allocation and dynamic respiration calculation, this embodiment constructs a crop sub-model that can respond to instantaneous environmental fluctuations and accurately capture growth and development trends throughout the entire growth cycle, providing a reliable biological kernel for greenhouse environment-yield co-simulation.

[0059] Furthermore, the energy consumption sub-model is used to quantify the electrical energy consumed by the greenhouse system in maintaining the target microclimate. As a key interface connecting environmental control strategies and economic benefit assessments, the accuracy of the energy consumption sub-model directly determines the guiding value of the simulation results in actual production management. Unlike the aforementioned climate and crop sub-models, which focus on natural physical and biological processes, the energy consumption sub-model is essentially a mathematical description of the energy cost incurred by human regulatory actions. Specifically, the formula for calculating heating energy consumption is: ; In the formula, For the control ratio of the boiler, For boiler power, The greenhouse floor area For boiler efficiency; A key feature of this formula is the explicit inclusion of boiler thermal efficiency as a denominator. In practical engineering applications, the heat generated by the boiler is not equal to the energy consumed by the fuel. Heat loss due to pipeline transport, incomplete combustion, and thermal inertia during start-up and shutdown all lead to significant efficiency losses. If only the theoretical heat load is calculated while ignoring the efficiency factor, the simulated energy consumption will be significantly lower than the actual billing data, thus misleading energy-saving decisions. By introducing this factor, this embodiment maps ideal thermodynamic requirements to real energy consumption, allowing simulation results to be directly integrated into the cost accounting system for greenhouse operation. It should be understood that although this embodiment uses a gas-fired hot water boiler as an example, in other embodiments, if other heating equipment such as heat pumps, ground source heat pumps, or electric heaters are used, only the boiler thermal efficiency needs to be considered. The core logic remains the same: "theoretical load divided by system energy efficiency," even if replaced with the corresponding coefficient of performance (COP) or electrothermal conversion efficiency. Both fall within the scope of protection of this invention.

[0060] The formula for calculating cooling energy consumption is: ; In the formula, For cooling load per unit area, This refers to the rated power of the fan-evaporative cooling pad system.

[0061] It is important to note that this embodiment only considers the mechanical power consumption of the fan-evaporative cooling system when calculating cooling energy consumption, intentionally ignoring the energy consumption of transient low-power devices such as ventilation window motors and sunshade motors. This boundary condition is based on the actual distribution characteristics of greenhouse energy consumption structure: during the high-temperature season, the fan-evaporative cooling system often operates continuously at high frequency for extended periods, accounting for over 90% of the total cooling energy consumption; in contrast, the roof windows and sunshades operate at low frequency, with short single-cycle durations and relatively low power, and their energy consumption contribution can be considered a secondary factor on a macroscopic time scale. More importantly, the frequent start-stop of these actuators introduces a large amount of high-frequency impulse noise into the energy consumption curve. Incorporating this into the integral calculation of the ODE model would not only significantly increase the stiffness of the numerical solution but may also force a reduction in the integration step size, thereby decreasing simulation efficiency. Therefore, focusing on the main energy-consuming devices for modeling ensures both the engineering accuracy of energy consumption prediction and the numerical stability of the hybrid model in the continuous time domain.

[0062] Furthermore, this embodiment uses a glass-enclosed tomato cultivation greenhouse as the simulation object, and constructs a simulation architecture for the multi-field coupling mechanism of environment, crop, and energy consumption. The system coupling topology is as follows: Figure 2 As shown, the model variables and the material and energy transfer links are divided according to the law of conservation of energy and the law of conservation of water vapor and carbon dioxide.

[0063] like Figure 2 As shown, the model variables are divided into three categories: 1. Circular icons represent system input variables, among which the outdoor meteorological input parameters include total outdoor solar irradiance. (i.e., global solar irradiance), equivalent sky radiation temperature Outdoor dry bulb temperature Outdoor water vapor pressure Outdoor carbon dioxide concentration The greenhouse bottom is equipped with equipment control inputs: boiler control signals. (i.e., boiler heating control), wet curtain control signal (i.e., evaporative cooling pad control), roof window opening and closing control (i.e., top window control), sunshade control signal (i.e., the start and stop control of the sunshade curtain), the four types of parameters are the control inputs that can be actively adjusted manually in this invention. 2. The solid-line box represents the state variables to be solved in the model, including the outer surface temperature of the greenhouse enclosure. (i.e., the outer surface temperature of the covering layer) and the inner surface temperature of the enclosure. (i.e., the inner surface temperature of the covering layer), the temperature of the sunshade component. Air temperature at the top of the greenhouse (i.e., the air temperature at the top of the greenhouse) and the air temperature inside the main body of the greenhouse. Crop canopy temperature Temperature of wet curtain packing (i.e., evaporative cooling pad temperature), heating pipe temperature (i.e., the temperature of the underfloor heating pipes) and the temperature of the greenhouse floor. (i.e., greenhouse ground temperature), multi-layer soil temperature ~ and the boundary temperature of the soil subsurface Supporting zoned water vapor pressure distribution: water vapor pressure distribution inside the enclosure Water vapor partial pressure of sunshade curtains Canopy water vapor partial pressure Indoor air water vapor partial pressure Water vapor partial pressure in the upper part of the greenhouse Zoned carbon dioxide concentration: Indoor main carbon dioxide concentration Carbon dioxide concentration in the upper part of the greenhouse 3. The dashed boxes represent semi-state variables, including... , , These variables are calculated in real time from the corresponding state variables and do not participate in the independent integration of ordinary differential equations.

[0064] Six different colored arrows are used to distinguish the energy and material transport paths within the system: blue arrows represent the conduction-convection sensible heat exchange pathway (i.e., sensible heat flux), realizing sensible heat transfer between the enclosure structure, air medium, crop plants, ground, and soil; yellow arrows represent the long-wave radiation heat exchange pathway, characterizing the exchange of infrared long-wave radiation energy between the various solid components of the greenhouse and the sky environment; reddish-brown arrows represent the short-wave solar radiation pathway (i.e., short-wave heat exchange), where incident solar short-wave radiation penetrates the greenhouse enclosure structure and then flows through the enclosed surface, shading curtains, crop canopy, and other components. The greenhouse floor absorbs heat layer by layer; the green arrows represent the latent heat exchange pathway (i.e., latent heat exchange), characterizing the latent heat migration accompanying crop transpiration, condensation on the inner wall of the enclosure, and water evaporation from the wet curtain; the magenta arrows represent the water vapor mass transport pathway (i.e., water vapor flux), realizing the diffusion of water vapor from high partial pressure areas to low partial pressure areas and the material transfer through ventilation between different greenhouse sections; the purple arrows represent the carbon dioxide gas transport pathway (i.e., carbon dioxide flux), depicting the carbon consumption of crop photosynthesis, spatial gas diffusion, carbon exchange between ventilation and the outdoor atmosphere.

[0065] Based on the aforementioned coupled topology, three sub-models are constructed: a climate sub-model, a crop sub-model, and an energy consumption sub-model. The climate sub-model solves for the internal environmental parameters of the greenhouse based on energy balance, water vapor exchange, and ventilation heat exchange relationships. The crop sub-model quantifies the accumulation of crop dry matter and yield generation patterns by combining canopy temperature, CO2 concentration, and photosynthetically active radiation. The energy consumption sub-model calculates the operating energy consumption of heating and cooling equipment based on the operating conditions of four types of control equipment. This invention adopts a two-stage hybrid modeling scheme: first, the parameters of the pure mechanistic model are calibrated; then, a hypernetic neural network is introduced to fit the nonlinear residuals of the system that the mechanistic model cannot accurately represent. The derivatives of the physical terms and the derivatives of the residual terms obtained from the mechanistic model are superimposed and input into an ordinary differential equation solver for integration. Simultaneously, three types of simulation results are output: greenhouse microclimate data, tomato yield data, and equipment energy consumption data, completing the integrated simulation calculation of the entire greenhouse system.

[0066] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for joint simulation of greenhouse environment output and energy consumption based on a hybrid model, characterized in that, Includes the following steps: Step S1: Obtain multi-source time-series input data of the greenhouse system, including time-varying external weather data and equipment control action data; Step S2: Calculate the physical derivatives of the greenhouse system based on the multi-source time-series input data using the physical model; Step S3: Utilize a neural network to dynamically generate network weights based on the time-varying external weather data, and calculate the derivative of the residual term based on the prior prediction output of the multi-source time-series input data and the physical model. Step S4: Combine the physical term derivative and the residual term derivative to obtain the total state derivative; Step S5: Input the total state derivative into the ordinary differential equation solver to perform continuous-time integration, and output the simulation results of the greenhouse system. The simulation results include indoor microclimate data, crop yield data, and equipment energy consumption data.

2. The method according to claim 1, characterized in that, The neural network adopts a supernetwork architecture that dynamically generates weights based on external time-varying conditions; The supernetwork architecture replaces static network weights with low-dimensional matrix products dynamically generated from the time-varying external weather data, so as to adaptively modulate the coupling relationship between the state variables of the greenhouse system according to environmental disturbances.

3. The method according to claim 1, characterized in that, The hybrid model consisting of the physical model and the neural network is obtained through progressive two-stage training paradigm. In the first phase, only physical models are used to correct uncertain parameters in order to anchor a set of parameters that conforms to the current greenhouse structure and crop characteristics; In the second stage, physical models and neural networks are jointly trained.

4. The method according to claim 3, characterized in that, The joint training process includes: The hybrid model is forward propagated based on a differentiable integrator, and numerical integration is performed in the time domain to solve for the continuous predicted trajectory. By utilizing a dynamic weighting mechanism based on the Softmax function, the multi-objective loss between the continuous predicted trajectory and the actual observation data is adaptively balanced; A multi-dataset polling update strategy is introduced to maintain the continuity of the time series, and a cosine annealing learning rate decay strategy is combined for gradient backpropagation and parameter optimization.

5. The method according to claim 1, characterized in that, The physical model consists of a climate sub-model, a crop sub-model, and an energy consumption sub-model. The physical model is formalized as a set of first-order continuous-time ordinary differential equations, driven by outdoor meteorological parameters, control action parameters, greenhouse design parameters, and crop physiological parameters, to describe the dynamic changes of the greenhouse system.

6. The method according to claim 5, characterized in that, The climate sub-model explicitly introduces the evaporative cooling pad temperature as an input variable to calculate the dynamic cooling efficiency of the fan-evaporative cooling pad system. The formula for calculating sensible heat flux in the climate sub-model is as follows: ； In the formula, For sensible heat flux, The ventilation flux generated by the fan-evaporative cooling pad system. air density; The specific heat capacity of air at constant pressure. Outdoor dry-bulb temperature, The latent heat of vaporization of water, For dynamic cooling efficiency, This refers to the moisture content of the evaporative cooling pad. The humidity content of outdoor air; The formula for calculating the steam flux under dynamic changes in indoor humidity is: ； In the formula, For steam flux, For the efficiency of the wet curtain; Ventilation flux of the fan-evaporative cooling pad system The calculation formula is: ； In the formula, The opening degree of the fan-evaporative cooling pad system; The ventilation capacity of the evaporative cooling pad; The greenhouse floor area; According to the equation for moist air in an adiabatic saturated process Represented as: ； In the formula, Represents the outdoor dry-bulb temperature. It serves as the thermodynamic wet-bulb temperature. The saturated moisture content of the evaporative cooling pad is given by the following formula: ； In the formula, The saturated vapor pressure at the wet curtain temperature. Absolute atmospheric pressure; The dynamic cooling efficiency The calculation formula is: ； In the formula, This refers to the temperature of the evaporative cooling pad. It is the thermodynamic wet-bulb temperature.

7. The method according to claim 5, characterized in that, The crop sub-model is based on the source-sink mechanism and simulates the process of carbohydrates from production and storage to conversion into crop organ dry weight. The formulas for calculating carbohydrate allocation fluxes in fruits, leaves, and stems in the crop sub-model are as follows: ； ； ； In the formula, As a carbohydrate deficiency inhibitor, As a factor that inhibits instantaneous temperature discomfort, It is an inhibitory factor for discomfort caused by 24-hour average temperature. It is an inhibitory factor that is unsuitable for crop development stages. The effect of temperature on carbohydrate flux in fruit. , , These are the potential growth rate coefficients for fruits, leaves, and stems, respectively.

8. The method according to claim 7, characterized in that, The calculation formulas for maintenance respiration and growth respiration in the crop sub-model are as follows: The formula for maintaining respiration is: ； In the formula, The weight of carbohydrates in plant organs. The maintenance respiration coefficient of plant organs. The effect of temperature on maintaining respiration value, The average canopy temperature over 24 hours. To maintain the respiratory regression coefficient, Relative growth rate; This represents the relative growth rates of three different types of organic matter. relative growth rate The calculation formula is: ； In the formula, For the maximum relative growth rate, This is the cumulative amount of daily radiation. It is the Michaelis constant; The formula for calculating growth and respiration is: ； In the formula, This represents the growth and respiration coefficient of plant organs; For the first The carbohydrate buffer capacity of plant organs.

9. The method according to claim 5, characterized in that, The energy consumption sub-model includes heating energy consumption and cooling energy consumption, which are used to quantify the electrical energy consumed by the greenhouse system in maintaining the target microclimate. The formula for calculating heating energy consumption is: ； In the formula, For the control ratio of the boiler, For boiler power, The greenhouse floor area For boiler efficiency; The formula for calculating cooling energy consumption is: ； In the formula, For cooling load per unit area, This refers to the rated power of the fan-evaporative cooling pad system.

10. The method according to claim 1, characterized in that, After acquiring the multi-source time-series input data of the greenhouse system, the following is also included: Data outliers are removed based on the boundary conditions of each parameter. The SAITS algorithm is used to impute missing data in order to obtain continuous and complete time series data.

Images