CFD simulation and measurement and control analysis of the ambient temperature field of agricultural greenhouses

In recent years, the application of Computational Fluid Dynamics (CFD) technology in greenhouse environmental control has received increasing attention. This numerical simulation-based approach can accurately simulate airflow, temperature distribution, and humidity changes inside greenhouses, providing a scientific basis for microclimate prediction and management. In the southeastern Yunnan region, where climatic conditions are complex and variable, regulating the greenhouse environment faces many challenges[19]. Against this backdrop, CFD technology provides theoretical support for optimizing the internal environment of greenhouses[5]. For instance, by simulating air circulation and temperature distribution within the greenhouse, researchers have proposed ventilation strategies and temperature control measures suited to different seasons and climate conditions, significantly improving energy efficiency and crop yields[32].

However, there are also limitations in the practical application of CFD. The accuracy of simulations is highly dependent on the precise setting of initial conditions and boundary conditions, which are often difficult to obtain in southeastern Yunnan due to the diversity of natural conditions and the complexity of data collection[16]. Additionally, CFD involves extensive and complex computations, which are time-consuming and may not respond quickly enough to the rapidly changing climatic conditions in southeastern Yunnan. While some scholars have introduced parallel computing or GPU-based acceleration techniques to improve simulation speed, these improvements have achieved some success in enhancing computational efficiency, but the real-time response capability still falls short of ideal levels[34].

When dealing with multivariable and multiscale meteorological data, CFD simulations are also prone to being affected by data redundancy and noise, which limits the accuracy and reliability of the simulation results[1]. Given the complex climatic background in southeastern Yunnan, further improving the accuracy and real-time performance of CFD models has become an urgent issue for researchers[30]. The latest research trends indicate that combining CFD technology with other methods, such as intelligent optimization algorithms or machine learning models, can enhance its predictive capability and response speed in complex environmental conditions[3][21]. These new approaches provide valuable insights into optimizing the application of CFD in greenhouse environmental control in southeastern Yunnan.

While the potential of CFD technology in greenhouse environmental control is immense, many challenges remain to be overcome in practice, especially in southeastern Yunnan. Therefore, this paper aims to explore pathways to combine CFD technology with other advanced methods in order to achieve more efficient and accurate greenhouse microclimate prediction and management.

With the rapid development of artificial intelligence, the application of deep learning in agricultural environment prediction has gradually gained attention. Deep learning models have the ability to adaptively learn from large amounts of data and capture complex nonlinear relationships, making them highly effective in data analysis and prediction in agricultural environments[17]. In greenhouse management, environmental parameters such as temperature, humidity, and carbon dioxide concentration are crucial for crop growth[15]. The introduction of deep learning offers new possibilities for the accurate prediction of these complex variables.

CNN as an important model in deep learning, have demonstrated unique advantages in agricultural data analysis. CNNs can effectively extract spatial features from multidimensional environmental data, capturing local variation patterns of environmental parameters such as temperature and humidity inside greenhouses[29]. Studies have shown that using CNNs to analyze sensor data from greenhouses can significantly improve the accuracy of environmental parameter predictions and provide more reliable data support for further environmental control[23]. The feature extraction capability of CNNs enables them to handle data from different types of sensors and integrate multi-source information, contributing to a more comprehensive understanding of microclimate changes within the greenhouse.

TCN excel at processing time-series data in agricultural environments. TCNs can capture long-term dependencies, making them suitable for predicting dynamic processes such as temperature and humidity changes. Compared to traditional RNNs, TCNs are more stable and efficient in learning time-series data, particularly excelling in long-term prediction tasks[4]. In the southeastern Yunnan region, TCN models can accurately predict future climate change trends based on historical data, which is crucial for timely adjustments to greenhouse environmental control strategies[8].

The characteristics of deep learning models give them broad application prospects in agricultural environment prediction. However, single deep learning models still have limitations when dealing with complex greenhouse environments, especially in predicting the coupling effects of multiple variables and handling nonlinear complex relationships[13][36]. Therefore, combining deep learning with other advanced methods has become a new research trend. This integrated approach not only addresses the limitations of traditional CFD simulations but also enhances their practical application in greenhouse environmental control in southeastern Yunnan, improving the overall system’s intelligence and response speed.

In recent years, the integration of deep learning techniques and CFD simulations has garnered increasing attention in the agricultural sector. This approach aims to leverage the powerful feature extraction and pattern recognition capabilities of deep learning models to address the limitations of CFD in processing complex data and real-time predictions[7]. By utilizing deep learning models for pre-processing and dynamic analysis of environmental data, more accurate input conditions can be provided to CFD, thereby enhancing the precision and efficiency of CFD simulations[26].

In greenhouse environmental control, CNN can be employed to analyze vast amounts of sensor data and extract critical spatial features, which provide more reliable boundary conditions and initial parameter settings for CFD simulations. For instance, CNNs can rapidly identify temperature gradients and humidity distribution patterns in different regions of the greenhouse and integrate this information into CFD models[10]. This integration optimizes the simulation results of airflow and heat distribution within the greenhouse, significantly reducing the processing time for complex environmental data in traditional CFD models while improving simulation accuracy.

On the other hand, Temporal Convolutional Networks (TCN) play a crucial role in dynamic prediction for greenhouse environments. TCNs can deeply analyze and forecast time-series data, such as predicting the future trends of temperature and humidity variations within the greenhouse[2][37]. By incorporating TCN prediction results as dynamic inputs into the CFD model, real-time simulation adjustments can be achieved, enabling greenhouse control systems to respond more flexibly to climate changes[35]. This integrated approach not only improves the real-time responsiveness of CFD simulations but also enhances the adaptability and effectiveness of greenhouse environmental control strategies.

In such research, some scholars have also explored the incorporation of optimization algorithms into the framework combining deep learning and CFD to further enhance the overall system performance[9]. By introducing intelligent algorithms such as Particle Swarm Optimization (PSO), control strategies and parameter settings can be dynamically optimized according to specific environmental requirements of the greenhouse[24]. In practical applications, the PSO algorithm can utilize deep learning models and CFD simulation results to rapidly search for optimal greenhouse management solutions, ensuring an ideal balance between energy efficiency and crop yield[18].

These explorations demonstrate that the integration of deep learning and CFD offers new possibilities for greenhouse microclimate control. By combining the strengths of both approaches, the accuracy and efficiency of greenhouse environment simulations can be significantly improved, providing more intelligent control solutions, especially in regions such as southeastern Yunnan, where the climate is complex and variable. The application potential of this method is vast, laying a solid foundation for the future development of smart agriculture.

In the overall DeepCFD-OptNet model, the Convolutional Neural Network (CNN) serves as the first core module, primarily used to extract key spatial features from various sensor data within the greenhouse (such as temperature, humidity, CO2 concentration, light intensity, etc.). The CNN identifies and extracts feature patterns from input data through convolution operations, making it particularly suited for handling multidimensional environmental data[20][25]. In this model, the CNN is responsible for extracting the most important feature information from complex multimodal data, which is crucial for the greenhouse environment simulation and control, providing more accurate initial conditions and boundary parameters for the CFD model.

The CNN module extracts higher-order features from the data by stacking multiple layers of convolution, pooling, and fully connected layers[31]. In the convolution layer, convolutional filters scan the local receptive field of the input data to identify local patterns or features in the data. (1) Zi,jl=∑m,n Xx+m,j+nl−1·Km,nl+bl where Zi,jl represents the output at position i,j in the l-th convolutional layer, X^l−1 is the input feature map from the previous layer, K^l is the convolution kernel in the l-th layer, and b^l is the bias term. Through the convolution operation, the CNN automatically learns different levels of data features, extracting spatial variation patterns within the greenhouse environment.

After processing through the activation function (ReLU), the output of the convolutional layer is passed through a pooling layer, which performs down-sampling to reduce the dimensionality of the feature map while retaining key features. The pooling formula is as follows: (2) Pi,jl=max(Xi′,j′l),wherei′∈[i,i+k),j′∈[j,j+k) where Pi,jl represents the output of the pooling layer, and k is the size of the pooling window. Pooling reduces data dimensionality while preserving important features.

In the overall model, the output of this module serves as input to the CFD model, providing crucial information about temperature gradients, humidity distribution, and other important aspects of different regions within the greenhouse. The spatial features extracted in this way significantly improve the initial accuracy of the CFD simulation. Additionally, these features provide a more accurate input foundation for the subsequent Temporal Convolutional Network (TCN), further enhancing the model’s dynamic prediction capabilities for the greenhouse environment. The output of the CNN module, in coordination with other modules, ensures the efficiency and reliability of the entire intelligent control system.

In the overall model, the Temporal Convolutional Network (TCN) serves as the second core module, primarily used for handling the time-series data within the greenhouse, such as the dynamic changes of temperature, humidity, and other parameters. The TCN captures long-term dependencies in the data through one-dimensional convolutions and dilated convolutions, making it highly effective in time-series prediction tasks[14][33]. In this model, historical data is used to predict the future changes in the greenhouse environment, providing a basis for the dynamic adjustments of the CFD model.

The key feature of this module lies in its convolutional operation, which can process time-series data in parallel, without relying on the sequential order. This avoids the gradient vanishing and exploding problems commonly encountered in RNNs when modeling long-term dependencies. In greenhouse environment control, the TCN model performs multi-layer convolutions on the input data, extending the receptive field through dilated convolutions to capture long-distance temporal dependencies. (3) y(t)=∑i=0k−1x(t−r·i)·w(i) where y(t) represents the current value of the output sequence, x(t−r·i) is the input sequence, w(i) is the convolution kernel, and r is the dilation rate. By increasing the value of r, the TCN can capture longer-term dependencies with relatively fewer layers.

To capture temporal dependencies effectively, the TCN model uses a series of convolutions with varying dilation rates, leading to an expanded receptive field: (4) y(t)=∑i=0k−1x(t−ri)·w(i)

Here, rⁱ indicates an increasing dilation rate as i increases, allowing the TCN to capture dependencies over a longer period. The dilation factor r allows the network to control the size of the receptive field. The relationship between the dilation rate and the receptive field can be expressed as: (5) Receptive Field=(k−1)·ri+1 where k is the kernel size and r is the dilation factor. This formula shows how the receptive field grows exponentially with increasing dilation rates.

In terms of multi-layer convolution, the TCN can stack several layers to learn increasingly abstract features of the time-series data. The output of each layer is fed as the input to the next layer, and the final output is calculated as: (6) yfinal=∑i=0L−1(∑j=0k−1x(t−rj·i)·wj(i)) where L is the number of layers, w_j(i) and represents the convolution kernels at each layer. This formulation highlights how the model applies multiple convolutional layers, with each layer contributing to the final prediction.

The loss function used to train the TCN can be formulated as a mean squared error (MSE) between the predicted and actual values, which is defined as: (7) L=1N∑t=1N(ypred(t)−ytrue(t))2 where N is the total number of time steps, y_pred(t) is the predicted value, and y_true(t) is the true value of the time-series data. The loss function is minimized during training to adjust the weights of the convolution kernels and improve the model’s prediction accuracy.

In this way, the module can efficiently predict future changes in environmental parameters within the greenhouse, such as forecasting temperature and humidity fluctuations for the next few hours. These prediction results provide more accurate and dynamic input data for the CFD simulation, enabling the CFD model to respond in real-time to rapid changes in both internal and external greenhouse environments, improving the model’s dynamic adaptability and prediction accuracy.

Figure 2.

Multi-group TCN network structure after convolution.

Overall, the TCN works in close collaboration with the CNN module. The spatial features extracted by the CNN provide a solid foundation for the TCN’s time-series modeling, allowing the TCN to more accurately analyze and predict the temporal dynamics of multidimensional data. This combination enables the model to achieve more efficient dynamic control and prediction in complex greenhouse environments, ensuring stability and optimization of the internal conditions. The output of the TCN module also integrates with the PSO optimization module, providing more accurate prediction data to optimize greenhouse control strategies, further enhancing the overall model′s intelligent regulation capabilities.

The Particle Swarm Optimization (PSO) algorithm, serving as the optimization mechanism of the entire model, is mainly used to optimize greenhouse environmental control strategies[11]. Based on swarm intelligence, the PSO algorithm simulates cooperation and information sharing among individuals in a group to find the optimal solution to complex problems. In this model, by utilizing the output results from the previous sections and CFD simulation data, it dynamically adjusts the greenhouse control parameters (such as the settings for ventilation, heating, and cooling systems) to ensure the best crop growth environment with minimal energy consumption.

In the overall model, the PSO algorithm is closely integrated with the deep learning modules and CFD simulation module. The feature data and prediction results provided by the CNN and TCN offer accurate reference points for the initial optimization in the PSO algorithm, while the CFD model provides real-time feedback on the environmental state during PSO iterations, making the optimization process more targeted and efficient. This synergy allows the model to quickly respond to environmental changes and intelligently adjust greenhouse control strategies, significantly enhancing the level of automation and economic efficiency in greenhouse management. (8) vit+1=w·vit+c1·r1·(pibest−xit)+c2·r2·(gbest−xit) (9) xit+1=xit+vit+1 where vit+1 and xit+1 represent the velocity and position of particle i at iteration t + 1, w is the inertia weight that controls the search range, c₁ and c₂ are learning factors, r₁ and r₂) are random numbers, pibest is the historical best position of particle i, and g^best is the global best position.

To further enhance the optimization process, the PSO algorithm dynamically incorporates feedback from the CFD module to fine-tune the particle updates. By integrating the environmental state data and energy consumption constraints, the model adjusts the movement of particles to achieve optimal control strategies. (10) Fitnessi(t)=α·Growthi(t)−β·Energyi(t)

where Fitnessi(t) represents the fitness value of particle i at iteration t, Growthi(t) is the predicted crop growth rate, Energyi(t) is the energy consumption, and α , β are weight coefficients balancing growth and energy considerations. (11) w=wmax−(wmax−wmin)·ttmax

where w is the inertia weight, w_max and w_min are its upper and lower bounds, t is the current iteration, and t_maxis the maximum number of iterations. This ensures a gradual reduction in exploration over time, enhancing convergence. (12) Δvi(t+1)=γ·∂Fitnessi(t)∂xi(t)+δ·∂CFDfeedback∂xi(t)

where Δv_i^(t+1)refines the velocity update based on gradients of fitness and CFD feedback, γ and δ are scaling factors for these contributions. (13) CFDfeedback=∑k=1N(∂2CFDk∂xk2+η·Thermalk)

where ^CFD_feedback aggregates thermal and structural effects from CFD simulations, providing a realtime adjustment to PSO particle movements. (14) xoptimal=arg minx(∑i=1PFitnessi(t)−λ·Constraint(x))

where x_optimal is the optimized control parameter set, Constraint(x) represents energy and environmental limitations, and λ is a penalty factor for constraint violations.

In greenhouse environmental control, the PSO algorithm continuously optimizes various parameters of the greenhouse control system through this process, achieving an optimal multi-objective control strategy that minimizes energy consumption and operational costs. This multi-level optimization design not only enables efficient environmental monitoring and prediction but also further optimizes resource allocation and control strategies within the greenhouse. It provides stronger adaptability and flexibility, capable of meeting different greenhouse management needs, and ensuring a continuously stable optimal growth environment for crops under the complex and variable climate conditions of southeastern Yunnan.

In this study, to accurately simulate and predict microclimate changes within greenhouses in the southeastern Yunnan region, we selected datasets from multiple sources for analysis and modeling. These data primarily include meteorological data from the China Meteorological Data Service Center and environmental monitoring data from inside the greenhouse, covering key environmental parameters such as temperature, humidity, light intensity, and CO2 concentration. This comprehensive dataset supports the training and validation of the model.

Historical meteorological data for the southeastern Yunnan region were obtained from the China Meteorological Data Service Center. These publicly available data come from ground meteorological stations operated by the China Meteorological Administration and include daily averages of temperature, humidity, wind speed, rainfall, and solar radiation, reflecting weather variations across different seasons and climate conditions[22][31]. The data span five years, providing a comprehensive representation of the climate characteristics and trends in southeastern Yunnan. During the data preprocessing stage, we first cleaned and standardized the raw meteorological data, removing outliers and noise to ensure accuracy and consistency. These meteorological data were then matched and integrated with the internal environmental data of the greenhouse, laying the foundation for subsequent model training.

Figure 3.

A Schematic diagram of the greenhouse environment and the monitoring equipment. On the left is the crop planting area in the greenhouse (a), and on the right are the environmental data recorder (b) and temperature and humidity monitoring equipment (c) respectively. These devices are used to monitor key environmental parameters such as temperature, humidity and CO2 concentration in the greenhouse.

The internal environmental monitoring data of the greenhouse were collected in real-time through a network of sensors deployed in various key areas of the greenhouse, including around the crops, near ventilation openings, and close to heating and cooling systems. These sensors monitored environmental parameters such as temperature, humidity, CO2 concentration, and light intensity with an hourly data collection frequency, capturing the dynamic changes in the greenhouse environment. These data provided critical feature inputs for the CNN and TCN models, enabling the models to accurately extract and analyze the spatial and temporal characteristics within the greenhouse. Additionally, the sensor data were used to calibrate the initial conditions and boundary parameters of the CFD model, improving the accuracy of the simulation.

To verify the effectiveness of the intelligent greenhouse microclimate prediction method based on CFD modeling proposed in this paper, we designed a multi-level experimental setup, including the configuration of the experimental environment, model training parameters, and validation methods. These settings are intended to ensure the reliability of the experimental results and optimal performance of the model.

The experimental environment setup includes both hardware and software aspects. On the hardware side, we used a server equipped with a high-performance GPU (such as the NVIDIA Tesla V100) to accelerate the model training and simulation processes. This server has 128 GB of memory and 2 TB of SSD storage, which can handle the requirements of large-scale data processing and complex computations. On the software side, we used the Python programming language and its deep learning frameworks (such as TensorFlow and PyTorch) to implement model construction, training, and optimization. Additionally, Fluent software was used for CFD simulations to ensure the accuracy and computational efficiency of the simulations. The entire experimental process was deployed and managed using Docker containers to standardize the environment and facilitate reproducible experiments.

For the model training parameters, we configured the CNN module with a three-layer convolutional structure, where the size of each convolution kernel is 3x3, and the number of channels is set to 32, 64, and 128, respectively. The activation function used is ReLU. The Adam optimizer was applied for training, with an initial learning rate set to 0.001, and a batch size of 64. To prevent overfitting, a dropout strategy was adopted with a ratio of 0.5. For the TCN module, we set a three-layer dilated convolution structure, with dilation rates of 1, 2, and 4, and a convolution kernel size of 3. The same Adam optimizer was used, with an initial learning rate of 0.0005 and a batch size of 32. Both models used an early stopping strategy to avoid overfitting. For the PSO algorithm module, the initial particle number was set to 50, with the velocity and position of each particle randomly assigned based on historical data and environmental conditions. The inertia weight was set to 0.7, and the learning factors were set to 1.5 and 2.0, respectively. The number of iterations for the PSO algorithm was set to 1000, ensuring that the optimal greenhouse control parameters were found in the multi-dimensional search space. In each iteration, the PSO module dynamically adjusted the greenhouse control strategy based on the input data from the CNN and TCN modules and the feedback results from the CFD simulations, aiming to minimize energy consumption and optimize the crop growth environment.

In this study, multiple evaluation metrics were used to comprehensively assess the model’s performance, including Mean Squared Error (MSE) and Mean Absolute Error (MAE) to measure the prediction accuracy of the model; computation time and resource consumption to evaluate the model’s computational efficiency; and energy cost and greenhouse environmental stability indicators (such as temperature and humidity fluctuation ranges) to evaluate the effectiveness of the control strategy. The experimental results were compared with those of traditional CFD models and other commonly used deep learning models (such as RNN and LSTM) to validate the advantages of the proposed model. (15) MSE=1n∑i=1n∑j=1nwj·(yi,j−y^i,j)2

where n is the number of samples, m is the dimension of the predictions, y_i,j is the actual value of the i-th sample in the j-th dimension, y^i,j is the corresponding predicted value, and W_j is the weight coefficient for each dimension, used to adjust the impact of different dimensions in the error calculation. (16) MAE=1n∑i=1n∑j=1nwj·| yi,j−y^i,j | (17) Energy Cost =∑t=1T∑k=1KPk(t)·Δt·C(t) )

where T is the number of time intervals for the total runtime, K is the number of devices, P_k(t) represents the power (in kilowatts) of the k-th device during the time interval t, Δt is the length of the time interval, and C(t) is the electricity price during the time interval t (in currency/kilowatt-hour). This formula takes into account the device power, runtime, and dynamic electricity pricing, providing a more accurate reflection of the energy cost of greenhouse operations.

Figure 4 shows the temperature change of the intelligent greenhouse microclimate prediction model in different times and heights based on CFD modeling in southeast Yunnan. It can be observed that the temperature distribution inside the greenhouse shows a distinct gradient with time and height. In the morning (09:00 to 11:00), the temperature in the greenhouse is relatively uniform, and the temperature near the ground is relatively low, which is in line with the natural law of low external temperature at night and the upward diffusion of heat from the ground. From noon to early afternoon (12:00 to 14:00), the temperature in the greenhouse increased rapidly, especially at the central height (about 600 to 1400 mm) and the upper area (over 1400 mm). This trend is due to direct sunlight and heating of indoor air. The deep learning enhanced CFD model proposed in this study accurately captures this changing process and showing the good adaptability of the model to dynamic environmental changes. In the afternoon (15:00 to 16:00), the temperature in the greenhouse gradually leveled, especially in the height of 2200 mm, the temperature distribution is more uniform. This phenomenon suggests that during the day, the accumulation and dissipation of heat in the greenhouse are mainly concentrated in the middle and low regions, while the temperature in the upper area is more stable. Through the prediction of this model, it can accurately reflect the temperature changes in different height and time periods, which provides an effective basis for optimizing the environmental control strategy of temperature room.

Figure 4.

Trend plot of temperature over time and height.

Model in the multidimensional data (time, space), through the CNN extraction space features, TCN modeling time series changes, and the PSO in optimizing greenhouse environment control parameters, successfully improve the accuracy of the greenhouse internal temperature control and the stability of energy efficiency management. The experimental results further validate the application effect of the model under the complex climate conditions in southeast Yunnan, and provide theoretical and practical support for the further promotion of smart agriculture in the region.

Figure 5 shows the influence of different model components on the prediction accuracy of relative humidity in the greenhouse. The ablation experiments of various models, including the removal of CNN, TCN and PSO, and the performance of the complete model DeepCFD-OptNet proposed in this paper. The experimental results show that the combination of different model components has significant effects on the prediction accuracy. In the model with CNN removal (red dashed line), the prediction error for relative humidity was maximized, with an RMSE of 3.21. This suggests that the CNN has an important role in extracting the spatial features of the greenhouse environmental data. The model removing TCN (green solid line) performed slightly better, with an RMSE of 2.49, but still inferior to the full model. This suggests that TCN is essential for improving prediction accuracy in terms of time series modeling. The model removing PSO (blue dots) performed relatively better, with an RMSE of 1.99, indicating that although PSO has an obvious effect in optimizing the control parameters of the greenhouse environment, its effect is not the most central in the model. The full DeepCFD-OptNet model (purple dot line) has the lowest RMSE value of 1.44, showing the best predictive performance. This suggests that the full model incorporating CNN, TCN, and PSO has significant advantages when handling complex dynamic changes in the greenhouse environment. The experimental results verify the effectiveness and superiority of the DeepCFD-OptNet model in greenhouse environment control. The model can make full use of the advantages of deep learning and optimization algorithms to realize the accurate prediction and control of complex environmental conditions, and optimize the energy efficiency management of greenhouse and crop growth environment.

Figure 5.

Effect of different model components on relative humidity prediction in ablation experiments.

Figure 6 shows the comparison of the DeepCFD-OptNet model proposed here and the traditional CFD model on the greenhouse temperature prediction. The red solid line represents the actual measured temperature data, the blue solid line represents the prediction results of the model presented here, and the gray dashed line represents the prediction results of the traditional CFD model. As can be seen from the figure, the prediction results of the DeepCFD-OptNet model are closer to the actual temperature data, and the overall trend changes are in good agreement with the actual data.

Figure 6.

The comparison of DeepCFD-OptNet model and the traditional CFD model on temperature prediction.

In contrast, the traditional CFD models were more biased, especially during time periods of sharp temperature changes (e. g., evening and early morning). This indicates that the traditional CFD models have some limitations in dealing with the complex temperature dynamics in the greenhouse, which is difficult to accurately reflect the real changes in the microclimate. Based on the introduction of deep learning modules (CNN and TCN) for feature extraction and time series modeling, the model further optimizes the greenhouse control parameters through the particle swarm optimization algorithm (PSO), thus significantly improving the accuracy of temperature prediction. The experimental results validate the superior performance of DeepCFD-OptNet model in the context of greenhouse environmental control, and they are able to simulate and predict the temperature change in the greenhouse more accurately, which provides important technical support for optimizing greenhouse management and improving crop yield.