Multi-Parameter Prediction of Solar Greenhouse Environment Based on Multi-Source Data Fusion and Deep Learning

Abstract

In the process of agricultural production in solar greenhouses, the key to the healthy growth of greenhouse crops lies in accurately predicting environmental conditions. However, there are complex couplings and nonlinear relationships among greenhouse environmental parameters. This study independently developed a greenhouse environmental acquisition system to achieve a comprehensive method for the monitoring of the greenhouse environment. Additionally, it proposed a multi-parameter and multi-node environmental prediction model for solar greenhouses based on the Golden Jackal Optimization-Convolutional Neural Network-Bidirectional Gated Recurrent Unit-Self-Attention Mechanism (GCBS). The GCBS model successfully captures the complex nonlinear relationships in the greenhouse environment and accurately predicts changes in carbon dioxide concentration, air temperature and humidity, and soil temperature at different location nodes. To validate the performance of this model, we employed multiple evaluation metrics and conducted a comparative analysis with four baseline models. The results indicate that, while the GCBS model exhibits slightly higher computational time compared to the traditional Long Short-Term Memory (LSTM) network for time series prediction, it significantly outperforms the LSTM in terms of prediction accuracy for four key parameters, achieving improvements of 76.89%, 69.37%, 59.83%, and 56.72%, respectively, as measured by the Mean Absolute Error (MAE) metric.

1. Introduction

China is among the countries with the largest area of facility planting in the world. The country has various forms of planting facilities such as plastic greenhouses, glasshouses, and solar greenhouses. By the end of 2023, China’s facility agriculture area totaled approximately 268 million hectares, with solar greenhouses comprising a notable proportion. Roughly 90% of these solar greenhouses are located in northern China, with most individual greenhouses ranging in total area from 300 to 1200 square meters [1,2]. As essential facilities for agricultural production in northern China, solar greenhouses provide suitable growing conditions for crops throughout the year, ensuring year-round production of fruit and vegetable products even during colder seasons. They are a primary means to achieve precision and efficient agricultural production [3,4]. China’s solar greenhouses are of the fully passive type, relying solely on their uniquely designed structures to accumulate heat and support plant growth during the winter without any auxiliary heating systems [5]. However, managers often rely on past experience to manually operate greenhouse equipment to maintain indoor temperatures. This traditional management approach is prone to operational errors and delays in regulation, ultimately affecting the stability of greenhouse environmental parameters and the growth efficiency of crops.

In recent years, the accelerated advancement of agricultural modernization and the widespread application of Internet of Things (IoT) technology have positioned solar greenhouses at the forefront as vital tools for precision and high-efficiency agriculture. The real-time monitoring and precise prediction of internal environmental factors within these greenhouses have emerged as significant research hotspots. As the scale of greenhouses continues to expand and crop growth requirements become increasingly complex, the effective regulation of greenhouse environments and the prediction of greenhouse states have gained importance [6,7]. By harnessing IoT technology, precise greenhouse environmental data can be collected, providing a solid foundation for the development of high-accuracy environmental prediction models [8]. These models, incorporating both internal and external environmental factors, enable the anticipatory prediction of greenhouse state trends, facilitating the timely adjustment of irrigation, ventilation, shading, and other measures [9,10]. In turn, this creates the best growth environment for crops, promoting their healthy and efficient growth and development. The application of IoT in optimizing greenhouse environments and resource management is sustainable, as it significantly conserves resources and reduces energy consumption while effectively mitigating the impact on the environment and human health [11].

Greenhouse environment modeling, a crucial component of precision and smart agriculture, aims to accurately simulate and predict complex and dynamic greenhouse environments. Numerous scholars have actively explored and applied various modeling approaches, including mechanistic models, time-series models, and deep learning-based time-series models, each with unique strengths and challenges [12].

Among these modeling approaches, mechanistic models are constructed based on principles such as energy and mass conservation, establishing multiple energy balance equations. Du et al. developed a simplified dynamic model for climate-smart greenhouse (CSG) vegetation dependency that dynamically simulates changes in air and soil temperature over time [13]. This model considers the unsteady energy balance in greenhouse air, including solar radiation, convective heat transfer through cover materials, heat transfer between soil and air, heat transfer between plants and air, energy losses due to ventilation, and heat supplied by the heating system. Singh et al. constructed a microclimate model for cucumber greenhouses in their study that predicts the temperature of the air, crops, cultivation substrate, and plastic covering under natural conditions [14]. However, they did not consider the spatial distribution characteristics of temperature and humidity within the greenhouse. Xu et al. analyzed the greenhouse temperature distribution mechanism under dynamic solar radiation, considering both crop transpiration and optical effects and constructed a mathematical model of greenhouse energy conversion based on the principle of energy balance [15]. This study sheds light on the mechanism underlying the formation of microclimate inhomogeneity. However, due to the complexity of the environment, the research did not delve into other environmental variables within the greenhouse. As a typical nonlinear and strongly coupled complex system, the solar greenhouse poses significant challenges. While mechanistic models provide deep insights into the physical mechanisms of the greenhouse environment, their complex modeling processes and numerous parameter settings limit their widespread application in practical agricultural production. Furthermore, these models tend to overlook the spatial distribution characteristics of environmental variables within the greenhouse, potentially limiting prediction accuracy.

Due to the slow and periodic nature of changes in greenhouse environments, their time-series data exhibit specific trends and cyclical patterns. This characteristic makes time-series modeling a vital tool for studying the patterns of greenhouse environmental variations and predicting future changes [16]. Guillén et al. used LSTM and four months of temperature data to predict whether extreme temperatures would occur in the greenhouse within the next 24 h, achieving a prediction error of less than 0.8 degrees [17]. The model only considered temperature data, though other variables in the greenhouse may also have an impact on the temperature prediction effect. He et al. predicted the minimum temperature in greenhouses to ensure the normal growth of crops [18]. In comparing various models, it was found that the prediction capabilities of deep learning models were significantly superior to those of generalized machine learning models, especially the gated recurrent unit (GRU) model. However, adjusting parameters in the training process of time-series models is highly dependent on human experience, leading to poor model generalization and high uncertainty. When dealing with high-dimensional data, these models tend to converge slowly and are prone to falling into local optima [19]. To address the problems in the training process of time series models, researchers have begun exploring automated methods of parameter tuning to reduce the need for manual intervention and improve the efficiency and accuracy of model training [20].

With the rapid development of deep learning technologies, integrating deep learning algorithms with time series models has gradually emerged as a new trend in greenhouse environment prediction. Scholars are dedicated to incorporating intelligent algorithms into greenhouse modeling, leveraging vast amounts of environmental data for training to construct various models. These models precisely capture the intricate correlations among greenhouse environmental factors and make accurate predictions based on these correlations. For instance, Chen et al. proposed combining particle swarm optimization and model predictive control algorithms for greenhouse air temperature control [21]. This algorithm effectively tracks the set greenhouse air temperature trajectory under disturbance. Gao et al. proposed a solar greenhouse temperature prediction model that considers the time-series characteristics and temperature dynamics of the day-light greenhouse system [22]. By conducting parameter analysis on the nonlinear auto-regressive exogenous neural network, they established a nonlinear autoregressive exogenous (NARX) neural network model. This model demonstrates high accuracy and performance in temperature prediction. The aforementioned models have demonstrated the superiority of deep learning-based time-series models in greenhouse environment prediction. In particular, Yang et al. constructed a Feedforward Attention Mechanism-Long Short-Term Memory (FAM-LSTM) model for multi-step prediction of temperature and humidity in solar greenhouses. This model considers multiple environmental factors and provides an effective method for temperature and humidity prediction [23]. Deep learning-based time-series models not only accurately predict long-term trends in greenhouse environments but also delve deeply into the complex relationships among environmental factors, thereby contributing to improved production and resource utilization efficiency in greenhouse agriculture. Ultimately, these models support the development of smart and precision agriculture.

In the field of monitoring and predicting environmental factors in solar greenhouses, the data collected by IoT technologies is typically vast and requires large-scale computational processing [24]. Deep learning-based time-series models can efficiently utilize this massive amount of data for training and excel at handling the complex nonlinear relationships and dynamic characteristics present in solar greenhouse environments [25].

This study aims to achieve high-precision predictions of multiple factors in solar greenhouse environments. We innovatively introduce the Golden Jackal Optimization (GJO) algorithm, an inspired swarm intelligence optimization algorithm modeled on the foraging and hunting behaviors of golden jackals in nature [26]. By integrating the strengths of GJO, Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Self-Attention Mechanism (SAM), we propose a multi-parameter and multi-node environmental prediction model for solar greenhouses based on GCBS. This model leverages large-scale IoT data from greenhouses as the training foundation, deeply exploring and effectively capturing the intricate nonlinear relationships and dynamic evolution patterns among environmental parameters. Experimental results demonstrate that the model performs exceptionally well in predicting future trends of key environmental factors such as air temperature, humidity, CO₂ concentration, and soil temperature within solar greenhouses. Compared to traditional and existing algorithmic models, the GCBS model exhibits significant advantages in prediction accuracy and network performance, providing greenhouse managers with more precise environmental change prediction information. This, in turn, enables them to make more scientific and rational decisions, significantly contributing to the precise regulation of multiple environmental factors in solar greenhouses.

2. Materials and Methods

2.1. Platform Construction and Data Acquisition

Constructing an IoT data acquisition system that can accurately and in real-time capture greenhouse environmental information is crucial for accurate greenhouse environmental prediction. We independently developed the “Wireless Network Monitoring System for Greenhouse Environmental Parameters.” The entire system comprises multiple layers: the information perception layer, the acquisition node layer, the gate-way layer, the local area network wireless data link layer, and the IoT management platform layer. The overall system architecture is illustrated in Figure 1. This IoT data acquisition system integrates embedded systems, low-power wide-area remote radio (LoRa), and 5G technology to monitor the greenhouse environment in real time. This innovative combination overcomes the limitations of single technologies in terms of coverage, transmission rate, and power consumption, significantly enhancing the real-time performance and accuracy of data collection.

The system adopts a star topology structure and achieves the flexible deployment of wireless networks through an intelligent acquisition network composed of micro-acquisition nodes and gateways. We designed a hierarchical network communication protocol and distributed algorithm to optimize communication efficiency, reduce conflicts, and enhance the self-organization and fault-handling capabilities of the network. This design ensures real-time, stable, and reliable data transmission, enabling the system to maintain efficient operation even in complex and varied greenhouse environments. The system performs tasks including sensing, acquisition, processing, and transmission of data while supporting functions such as synchronous acquisition, asynchronous transmission, and remote monitoring.

The information perception layer collects environmental data from both inside and outside the greenhouse, including light intensity, air temperature and humidity, CO₂ concentration, and soil temperature. The acquisition node layer receives instructions from the gateway to collect environmental data and processes and encapsulates the data to ensure its accuracy and integrity. The gateway layer acts as the central hub for data transmission and facilitates the interaction between acquisition nodes and the IoT platform. The wireless data link layer of the local area network leverages the advantages of LoRa and 5G technologies to ensure the security and efficiency of data transmission, realizing the connection between the system and the remote IoT platform as well as the transmission of data and instructions. The IoT platform management layer parses, processes, and stores greenhouse data received, presenting it through a graphical interface for users to conveniently access real-time status and historical data on the greenhouse environment. This system successfully achieves real-time monitoring of the greenhouse environment, improving data continuity and scientific validity. It provides an effective and reliable research platform and crucial data support for studying spatial and temporal distribution differences within the greenhouse environment.

2.2. Overview of the Experimental Area

This study was conducted in the greenhouse of farmers in Jicun Village, Taigu County, Jinzhong City, Shanxi Province, China (112°74′ E, 37°49′ N), which is characterized by a typical temperate continental monsoon climate. The geographic location of the IoT greenhouse and the collection method are shown in Figure 2. During the experimental period (December 2021 to March 2022), the tomato plants in the greenhouse were in the flowering stage, with a height of approximately 1 m, providing an ideal test subject for studying the impact of greenhouse environments on crop growth.

The greenhouse is structurally designed with rationality, featuring a north-south orientation, a length of 100 m, and a width of 8 m. The rear insulating wall is a soil wall with a height of 4 m and a thickness of 1.4 m, while the east and west walls are 0.5 m thick, with bricks on the outside and soil on the inside. Irrigation is conducted through furrow irrigation, with a ridge width of 62 cm, facilitating crop growth and water management. The light-transmitting material is polyvinyl chloride film, while the insulation blanket is made of cotton curtain material with a waterproof silk cloth outer surface and fiber cotton filling. The load-bearing structure is designed as a double-skeleton cable-stayed steel frame. Apart from ventilation openings and insulation blankets, the greenhouse is not equipped with other complex control devices, providing an authentic experimental environment for studying the natural variations in greenhouse environments.

2.3. Sensor Layout Plan

The interior of the greenhouse has been meticulously planned, with the space evenly divided into multiple zones along the north-south direction, each equipped with a data acquisition node to ensure comprehensive and representative data collection. Additionally, taking into account the growth characteristics of tomato plants and the structural features of the greenhouse, air temperature and humidity sensors, as well as soil temperature sensors, have been strategically positioned both vertically and horizontally to comprehensively capture subtle differences in the greenhouse’s internal environment across these dimensions.

As illustrated in Figure 3, five data acquisition nodes are strategically positioned within the greenhouse, labeled sequentially from Node 1 to Node 5 from south to north, with each node separated by 20 m. Collectively, these nodes monitor crucial environmental parameters, encompassing light intensity, air temperature and humidity, carbon dioxide concentration, and soil temperature. To facilitate a deeper comprehension of the interaction between the interior and exterior environments of the greenhouse, an additional node, designated as Node 0, has been installed outside the greenhouse to collect data on external factors such as light intensity, air temperature and humidity, soil temperature, wind speed, and wind direction, facilitating comprehensive analysis in subsequent stages.

Considering the common practice of topping tomato plants during their growth cycle to manage their height, a canopy height of approximately 2 m is typically achieved. In our investigation of the vertical distribution characteristics of greenhouse environmental parameters, the maximum height for data acquisition was set at 2.5 m to ensure comprehensive coverage above the tomato plant canopy. The air temperature and humidity sensors were meticulously deployed at five distinct levels, positioned at heights of 0.5 m, 1.0 m, 1.5 m, 2.0 m, and 2.5 m above the ground surface. Additionally, soil temperature sensors were installed at three varying depths, specifically at 10 cm, 20 cm, and 30 cm below the surface, to enable comprehensive monitoring of soil temperature fluctuations.

2.4. Data Preprocessing

Utilizing the aforementioned IoT data collection system, the data is uploaded to the server with a sampling interval of 30 min. To establish a more accurate prediction model for the greenhouse environment, the average value of multiple sensors from each node is adopted as the model training data. Considering the example of January 2022, during the collection of CO₂ concentration data, some data was missing, attributed to sensor malfunctions, data transmission interruptions, or external interference.

To ensure the completeness and accuracy of our dataset, we employed a series of innovative methods to safeguard data quality. To address the issue of missing data, we integrated the k-means clustering algorithm to gain insights into data distribution patterns, leveraging a combined strategy of imputation using nearby values and mean filling to estimate and fill in missing values [27]. As depicted in Figure 4, the combined use of these two strategies takes into account both the temporal sequence characteristics of the data and ensures the rationality and accuracy of the filled values. Following the completion of the missing value imputation, we proceed with outlier detection within the dataset. By setting thresholds for each environmental variable, we identify and handle outliers, thereby enhancing the reliability of the dataset. Lastly, to eliminate the dimensional differences among various environmental variables, we applied the Min-Max normalization method to standardize all variables [28]. This step ensures data consistency and provides a more reliable data foundation for subsequent model training.

2.5. Analysis of Horizontal Distribution Differences in Key Environmental Parameters of the Greenhouse

After the preprocessing steps, we obtained a comprehensive dataset. Through an in-depth analysis of the data collected from multiple nodes, we were able to observe the variations and changing patterns of the internal environmental parameters within the greenhouse.

Figure 5 demonstrates the distribution of environmental parameters in a solar greenhouse over time. The environmental parameters within the solar greenhouse exhibit distinct regular patterns of variation, which can be categorized into four stages: before the closure of insulation facilities, after the closure of insulation facilities, after the activation of ventilation facilities, and after the reopening of insulation facilities.

Prior to the closure of insulation facilities, the greenhouse was in a nighttime state. During this period, crops engage in aerobic respiration, releasing significant amounts of water and carbon dioxide, resulting in a gradual increase in the concentration of carbon dioxide within the greenhouse to its peak value. The relative humidity of the air remains at a relatively high level. The internal ambient temperature of the greenhouse declines slowly but consistently remains above 10 °C, owing to the unique architectural structure of the greenhouse. The heat within the greenhouse during the night primarily originates from the heat storage capabilities of the soil and insulation walls, indicating that the solar greenhouse possesses excellent insulation and heat storage properties, which contribute to a relatively stable internal environment.

Around 9:00, with the insulation facilities closed and ample sunlight, photosynthesis commences in the crops, leading to a decrease in the CO₂ concentration and relative humidity within the greenhouse. The sun’s radiation elevates both the internal temperature and soil temperature of the greenhouse, further augmenting the heat within.

At approximately 11:00, the ventilation facilities are activated, enabling gas exchange between the interior and exterior of the greenhouse. Consequently, the CO₂ concentration rapidly diminishes to its lowest level. The humidity inside the greenhouse gradually converges with the ambient environmental level, with the relative humidity decreasing slowly, albeit slightly faster, in the vicinity of the ventilation openings. The internal temperature of the greenhouse reaches its peak around 13:00 after a brief decline, subsequently commencing a gradual descent. The soil temperature, meanwhile, continues to ascend from 11:00 to 17:00, ultimately reaching approximately 14.5 °C.

Around 17:30, the insulation facilities are reactivated and the ventilation facilities are closed as the sunlight gradually fades inside the greenhouse, initiating the insulation phase. The crops resume aerobic respiration, resulting in an incremental increase in CO₂ concentration, albeit at a reduced rate. The internal temperature and soil temperature of the greenhouse undergoes a gradual decline. In contrast to the temperature trends, the relative humidity swiftly rebounds and remains at a high level. These variations underscore the intricate coupling relationships among the various environmental parameters within the greenhouse, collectively sustaining an optimal growth environment that favors crop development [29].

2.6. Analysis of Vertical Distribution Differences in Key Environmental Parameters in the Greenhouse

Figure 6 illustrates the variation of environmental parameters within a solar greenhouse according to height. Figure 6a depicts the soil temperature changes at different depths. The soil temperature within 0.1 m experiences notable variations upon activation and deactivation of the insulation facilities, indicating a close correlation between soil temperature and indoor ambient temperature. The soil temperature declines to approximately 12.5 °C around 9:00 and peaks at approximately 14.4 °C around 16:00. In contrast, soil temperature at a depth of 0.3 m undergoes minimal fluctuations, remaining stable between 13.5 °C and 14.0 °C with less external influence and primarily governed by the stability of the subsurface temperature.

Figure 6b shows the vertical temperature variation. During the day, the internal temperature of the greenhouse exceeds the soil temperature, while the reverse is true at night. This phenomenon underscores the thermal exchange process between the soil and the interior of the greenhouse, particularly the heat dissipation from the soil and insulation walls during the night. After the insulation facilities are deactivated, the interior of the greenhouse exhibits a single-peak pattern influenced by solar radiation, which is closely tied to the heat distribution and lighting conditions within. The indoor temperature reaches its peak around 13:00. The temperature increases with height, with the maximum temperature at 0.5 m reaching approximately 22.0 °C and peaking at 35.0 °C at 2.5 m.

Figure 6c displays the vertical humidity variation. The air’s relative humidity remains high throughout the day at 0.5 m due to its proximity to the ground, where evaporation and transpiration from crops contribute to a persistent local humidity. When the insulation facilities are activated, the interior of the greenhouse maintains a high humidity level. Around 9:00, as the insulation facilities are deactivated and photosynthesis commences in the crops, humidity slightly decreases at 2.5 m and 2.0 m. Following the activation of the ventilation facilities, air exchange between the interior and exterior of the greenhouse removes excess moisture, initiating an overall decline in humidity. The overall humidity drops, with the minimum humidity at 1 m reaching approximately 90% and that at 2.5 m dropping to about 65%. This disparity arises from the varying degrees of impact that ventilation has on different heights, as relative humidity decreases with increasing height.

3. Model Theory

3.1. Golden Jackal Optimization Algorithm

The GJO algorithm was proposed by Nitish Chopra and Muhammad Mohsin Ansari [30]. It is a swarm intelligence optimization algorithm that mimics the cooperative hunting behavior of golden jackal swarms in nature. Based on the leadership of male and female jackals guiding the hunting and resting activities of the pack, the algorithm mainly consists of three stages: prey searching by the golden jackal population, prey encirclement, and prey attack. Compared to the Grey Wolf Optimizer (GWO), the GJO algorithm introduces the concept of male and female jackals, refining the leadership mechanism and thereby enhancing the flexibility and convergence speed of the search process. This refinement also imparts stronger local search capabilities to the GJO algorithm [31]. Similarly, in contrast to the Particle Swarm Optimization (PSO) algorithm [32], which primarily relies on the velocity and position updates among particles to locate the optimal solution, the GJO algorithm mimics the hunting strategies of golden jackals, enabling more efficient utilization of swarm intelligence. This approach enhances the algorithm’s global search capabilities and adaptability.

3.2. Convolutional Neural Networks

Convolutional neural networks excel at processing data with grid-like structures. Its main components include convolutional layers, pooling layers, and fully connected layers [33]. The convolutional layers are responsible for extracting spatial features from the input data and are able to effectively capture local features by sharing parameters at different locations. The pooling layers preserve important features while reducing spatial dimensions, helping to improve the model’s robustness and reduce overfitting. The fully connected layers map high-level features to target categories. Due to the large number of parameters, some neurons are randomly zeroed during training to reduce dependencies among neurons and enhance the model’s generalization ability. However, CNN cannot extract time series features, so BiGRU is introduced for learning temporal characteristics.

3.3. Bidirectional Gated Recurrent Unit Neural Network

The Gated Recurrent Unit (GRU) neural network is an improved architecture of the Recurrent Neural Network (RNN). The BiGRU neural network consists of forward and backward GRU units. Cho et al. proposed the GRU neural network based on LSTM, which simplifies the structure, reduces parameters, and makes the model lighter and faster in training speed [34,35]. In greenhouse environments, environmental conditions at consecutive moments are closely related and require deeper-level feature extraction. By combining forward and backward GRU units, BiGRU can fully utilize temporal dependencies and explore feature relationships to enhance model performance.

3.4. Self-Attention Mechanism

The attention mechanism is a resource allocation mechanism that mimics the human brain’s attention mechanism, and its core role is to enhance attention and weight allocation to different positions in the input sequence. While traditional attention mechanisms can enhance a model’s focus on crucial information to a certain extent, their attention allocation relies heavily on external information guidance [36]. The SAM is capable of capturing long-distance dependencies within the input sequence, automatically paying attention to information from different locations. When tackling large-scale data and constructing complex, large-sized models, the SAM demonstrates remarkable scalability and adaptability [37]. In contrast to traditional attention mechanisms, the SAM is capable of directly modeling the relationships among elements within the input sequence, thereby transcending the limitations of traditional attention mechanisms that solely focus on the interaction between input and output [38]. By introducing the SAM into the BiGRU model, we can dynamically calculate the interrelationships among elements in the input sequence, enabling weighted inputs for each time step. The outputs in the BiGRU network are updated and iteratively computed through the SAM to derive a better matrix of weighting parameters to highlight the influence of different features in the data.

3.5. GCBS Hybrid Network Model

The internal environment of solar greenhouses is influenced by various mutually independent time series features, including outdoor air temperature, outdoor air relative humidity, indoor temperature, indoor air relative humidity, indoor light intensity, and soil temperature. In order to integrate this feature information pertaining to the greenhouse, discern potential relationships and spatial patterns, and accurately forecast the spatial distribution of diverse environmental factors within solar greenhouses, we devised the GCBS multi-parameter and multi-point prediction model tailored specifically for solar greenhouse environments. The network architecture of the GCBS model is illustrated in Figure 7.

The overall model is divided into four parts: Firstly, the GJO optimization algorithm is used to iteratively select the optimal hyperparameters of the model, aiming to find the best combination of parameters to enhance the model’s performance and effectiveness. Secondly, CNN is employed to extract features from the data, eliminate potential instability factors, and enhance the model’s ability to express the input data. The third part involves BiGRU, which analyzes environmental data and extracts feature vectors, helping to enhance the model’s understanding of sequential data. The final part is the SAM mechanism, which assigns different weights to the feature vectors to highlight effective features and aid the model in utilizing feature information more effectively. Finally, the processed features are input into the model’s output layer through a fully connected layer to obtain the final prediction results.

3.6. Model Evaluation Metrics

The device environment used for training is as follows: the processor is an Intel i7-13700H (Intel, Santa Clara, CA, USA) with 16GB of memory, the GPU graphics card is an NVIDIA RTX4060 (NVIDIA, Santa Clara, CA, USA), the operating system is 64-bit Windows 11, and the programming software is Matlab R2023b. We employ several metrics as performance evaluation indicators, including mean absolute error (MAE), mean absolute percentage error (MAPE), mean squared error (MSE), coefficient of variation of root mean squared error (CV-RMSE), and coefficient of determination (R²) [39,40].

Among these, a lower MAE and MAPE indicate the high prediction accuracy of the model while being insensitive to outliers. A smaller MSE reflects a model with high prediction precision and sensitivity to outliers. CV-RMSE is used to assess the relative variation in the model’s prediction errors, with a lower value signifying smaller errors and better consistency during prediction. Lastly, an R² value closer to 1 indicates superior predictive performance of the model. The calculation formulas for these required indicators are as follows:

$$\mathrm{MAPE}={\displaystyle \frac{100\%}{\mathrm{n}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}\left|\frac{{\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}}}{{\mathrm{y}}_{\mathrm{i}}}\right|}$$

(1)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}})}^2}}$$

(2)

$$\mathrm{MAE}={\displaystyle \frac{1}{\mathrm{n}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}}\right|}$$

(3)

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}})}^2}}{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\overline{\mathrm{y}}}_{\mathrm{i}})}^2}}}$$

(4)

$$\mathrm{MSE}={\displaystyle \frac{1}{\mathrm{n}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}})}^2}$$

(5)

$$\mathrm{CV}-\mathrm{RMSE}={\displaystyle \frac{\mathrm{RMSE}}{{\overline{\mathrm{y}}}_{\mathrm{i}}}}\times 100\%$$

(6)

Herein, ${\mathrm{y}}_{\mathrm{i}}$ denotes the actual value, ${\overline{\mathrm{y}}}_{\mathrm{i}}$ signifies the mean of those actual values, ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ represents the forecasted value, and $\mathrm{n}$ stands for the total count of data points involved.

4. Results and Analysis

4.1. Model Optimization and Training

The experimental data were sampled from December 2021 to March 2022, with a sampling interval of 30 min, resulting in approximately 5700 data points for single-node environmental parameters. Subsequently, these data were partitioned into a training set and a test set, adhering to an 8:2 ratio. Eight environmental factors, including outdoor air temperature, outdoor air relative humidity, outdoor ground temperature, indoor light intensity, indoor air temperature, indoor air relative humidity, indoor CO₂ concentration, and indoor soil temperature, were selected as input parameters for the model. Additionally, four environmental factors, namely indoor air temperature, indoor air relative humidity, indoor CO₂ concentration, and indoor ground temperature were selected as the output parameters of the multi-parameter multipoint prediction model for the next hour. These outputs were used to assess the overall environmental conditions within the daylight greenhouse.

The parameter settings for the optimization training of the model are as follows: population size of 15 for the golden jackal optimizer; a learning rate search range of [10⁻⁴, 10⁻²]; a search range for the number of neurons in the hidden layer of [10, 100]; a search range for the key value pairs in the attention mechanism of [2, 50]; a regularization parameter search range of [10⁻⁴, 10⁻²]; and a maximum training epoch set to 150. During the training process, the GJO optimization algorithm is utilized to continuously adjust the learning rate, the number of BiGRU hidden neurons, the key-value pairs in the attention mechanism, and the regularization parameters of the model. Additionally, a time step of 5 is set to achieve better model performance and training effectiveness.

To thoroughly validate the effectiveness of the GCBS model, this study selected four other models for comparative experiments: LSTM, BiGRU, CNN-BiGRU, and CNN-BiGRU-SAM. During the experimental process, network parameters were repeatedly adjusted using a controlled variable approach, and the prediction accuracy was judged by comparing the MAPE values. Initially, the number of neurons in the BiGRU network was fixed at 20, and the number of CNN layers was gradually adjusted [41]. As shown in Figure 8a, appropriately increasing the number of CNN layers can enhance prediction accuracy. When the CNN was designed with two layers, the MAPE was relatively low. After determining the optimal number of CNN layers, the number of hidden neurons in the BiGRU was adjusted. According to Figure 8b, the model achieved the lowest MAPE when using 15 hidden layer units. After fixing the CNN and BiGRU layers, the learning rate of the network was further adjusted. As illustrated in Figure 8c, a learning rate of 0.001 resulted in a relatively low MAPE. Through this series of experiments and comparisons, we were able to determine the prediction accuracy of the CNN and BiGRU models under different network layer configurations and select the optimal number of layers for further model optimization and application.

The data collected in the solar greenhouse is primarily time-series data. Therefore, we have selected a CNN module that comprises two 1D convolutional layers, two max-pooling layers, and a fully connected layer. The first 1D convolutional layer is used to capture low-level features, while the second 1D convolutional layer is responsible for capturing more advanced features, which helps to reduce the risk of overfitting and enhance the model’s generalization ability. Employing max pooling for pooling operations can reduce the computational burden, improve the model’s robustness to positional changes, and assist in preserving important features while reducing data dimensionality, thereby enhancing the model’s efficiency and performance. The feature data extracted through the convolutional layers is converted into a one-dimensional vector and then processed through the fully connected layer, effectively extracting feature data from the convolutional layers.

After conducting comparative experiments and analysis, we integrated the CNN and BiGRU models to leverage temporal correlations and explore feature relationships, further enhancing prediction performance. By concatenating the CNN and BiGRU models, we can fully capitalize on the strengths of both models, thereby improving the accuracy and generalization capabilities of the prediction model. The CNN module effectively extracts spatial features from data through convolutional operations, while the BiGRU module possesses robust temporal modeling capabilities, capable of capturing long-term dependencies within time series. By combining these two models, we can better handle complex time-series data and adapt to varying greenhouse environments.

In predicting multiple environmental factors within a greenhouse, we further integrated the CNN-BiGRU model with the SAM to better capture the associations and importance of the data. The environmental factors within a greenhouse often exhibit complex temporal relationships, and the SAM can assist the model in capturing long-distance dependencies. Additionally, the environmental factors within a greenhouse are typically multi-dimensional. By incorporating the SAM, the model can dynamically model the associations and importance among different environmental factors, thereby improving the prediction performance for multiple environmental factors and better adapting to complex and variable environments.

After numerous iterations of parameter adjustments, we discovered that the models achieved optimal prediction performance with an initial learning rate of 0.001, 150 iterations, 15 hidden layer nodes in the BiGRU layer, and a batch size of 128. To ensure comparability across experiments, we utilized the same dataset, optimal network structure, and fundamental hyperparameters. This consistent setup facilitated a more accurate assessment of model performance differences and enabled effective model comparison and analysis.

4.2. Model Results and Analysis

To validate the predictive performance of the models, a comparative analysis was conducted using the internal air temperature, relative air humidity, CO₂ concentration, and soil temperature as the outcomes of the prediction tasks.

Within the greenhouse, fluctuations in CO₂ concentration were quite significant, resulting in notable differences in prediction accuracy among various models.

Table 1. CO₂ concentration performance prediction, based on different prediction models: Model 1 represents the LSTM model; Model 2 signifies the BiGRU model; Model 3 denotes the CNN-BiGRU model; Model 4 indicates the CNN-BiGRU-SAM model; and Model 5 signifies the GCBS model.

Forecasting Task		CO2 Concentration Prediction
Model Number		Model 1	Model 2	Model 3	Model 4	Model 5
Node 1	R2	0.897	0.946	0.952	0.976	0.986
	MAE/(mg/kg)	68.500	49.100	44.700	30.100	21.700
Node 2	R2	0.873	0.937	0.967	0.979	0.988
	MAE/(mg/kg)	100.700	67.200	48.200	37.900	23.900
Node 3	R2	0.895	0.933	0.947	0.965	0.986
	MAE/(mg/kg)	85.500	48.600	42.800	33.900	23.700
Node 4	R2	0.874	0.939	0.954	0.974	0.990
	MAE/(mg/kg)	112.480	58.900	48.800	37.700	19.500
Node 5	R2	0.887	0.931	0.965	0.980	0.990
	MAE/(mg/kg)	105.400	58.800	42.400	32.400	20.400
Average node	R2	0.885	0.937	0.957	0.975	0.988
	MAE/(mg/kg)	94.516	56.520	45.380	34.400	21.840

compares the prediction accuracy for CO₂ concentration in the test set, and Figure 9a illustrates the model prediction results. The results indicate that Model 5 exhibited the best performance in predicting CO₂ concentration, with an R² of 0.988 and an MAE of 21.84 mg/kg for the prediction set. Compared to Models 1, 2, 3, and 4, Model 5 improved R² by 11.61%, 5.42%, 3.24%, and 1.35%, respectively, and reduced MAE by 76.89%, 61.36%, 51.89%, and 36.51%, respectively.

Due to the characteristics of the greenhouse environment, air humidity usually remains at a high and relatively stable level with minimal fluctuations. A comparative analysis of the prediction accuracy of various models for relative air humidity in the test set is presented in

Table 2. Air humidity performance prediction, based on different prediction models.

Forecasting Task		Air Humidity Prediction
Model Number		Model 1	Model 2	Model 3	Model 4	Model 5
Node 1	R2	0.763	0.857	0.916	0.941	0.958
	MAE/(%)	1.300	0.880	0.740	0.750	0.460
Node 2	R2	0.801	0.873	0.903	0.932	0.951
	MAE/(%)	1.010	0.920	1.500	0.710	0.490
Node 3	R2	0.791	0.917	0.923	0.944	0.960
	MAE/(%)	2.940	1.080	1.050	0.880	0.560
Node 4	R2	0.773	0.858	0.921	0.939	0.960
	MAE/(%)	1.570	1.030	0.890	0.740	0.440
Node 5	R2	0.786	0.861	0.928	0.935	0.955
	MAE/(%)	1.310	1.160	0.830	0.790	0.540
Average node	R2	0.783	0.873	0.918	0.938	0.957
	MAE/(%)	1.626	1.014	1.002	0.774	0.498

, and the model prediction results are illustrated in Figure 9b. The results demonstrate that Model 5 exhibits the best performance in predicting relative air humidity, with an R² of 0.957 and an MAE of 0.498% for the prediction set. Compared to Models 1, 2, 3, and 4, Model 5 improves R² by 22.23%, 9.57%, 4.20%, and 1.98%, respectively, and reduces MAE by 69.37%, 50.89%, 50.30%, and 35.66%, respectively.

The temperature inside the greenhouse typically exhibits regular and periodic fluctuations. A comparative analysis of the prediction accuracy of various models for air temperature in the test set is presented in

Table 3. Air temperature performance prediction, based on different prediction models.

Forecasting Task		Air Temperature Prediction
Model Number		Model 1	Model 2	Model 3	Model 4	Model 5
Node 1	R2	0.885	0.930	0.949	0.959	0.969
	MAE/(°C)	1.350	1.070	0.800	0.720	0.580
Node 2	R2	0.868	0.911	0.944	0.953	0.963
	MAE/(°C)	1.450	1.230	0.880	0.730	0.610
Node 3	R2	0.877	0.922	0.947	0.958	0.972
	MAE/(°C)	1.570	1.170	0.910	0.770	0.570
Node 4	R2	0.878	0.927	0.944	0.956	0.969
	MAE/(°C)	1.360	1.000	0.800	0.710	0.540
Node 5	R2	0.888	0.928	0.957	0.963	0.972
	MAE/(°C)	1.290	1.050	0.750	0.700	0.520
Average node	R2	0.879	0.924	0.948	0.958	0.969
	MAE/(°C)	1.404	1.104	0.828	0.726	0.564

, and the model prediction results are illustrated in Figure 9c. The results indicate that Model 5 performs relatively well in predicting air temperature, with an R² of 0.969 and an MAE of 0.564 °C for the prediction set. Compared to Models 1, 2, 3, and 4, Model 5 improves R² by 10.21%, 4.92%, 2.19%, and 1.17%, respectively, and reduces MAE by 59.83%, 48.91%, 31.88%, and 22.31%, respectively.

The soil temperature exhibits a consistent trend with the ambient environmental temperature, displaying a distinct diurnal pattern. A comparative analysis of the prediction accuracy of various models for soil temperature in the test set is presented in

Table 4. Soil temperature performance prediction, based on different prediction models.

Forecasting Task		Soil Temperature Prediction
Model Number		Model 1	Model 2	Model 3	Model 4	Model 5
Node 1	R2	0.906	0.925	0.937	0.960	0.970
	MAE/(°C)	0.078	0.068	0.060	0.057	0.043
Node 2	R2	0.765	0.917	0.933	0.958	0.970
	MAE/(°C)	0.171	0.079	0.061	0.059	0.041
Node 3	R2	0.922	0.937	0.951	0.971	0.975
	MAE/(°C)	0.089	0.079	0.067	0.058	0.048
Node 4	R2	0.869	0.911	0.929	0.970	0.983
	MAE/(°C)	0.088	0.071	0.064	0.057	0.046
Node 5	R2	0.927	0.921	0.945	0.958	0.968
	MAE/(°C)	0.080	0.070	0.073	0.054	0.041
Average node	R2	0.878	0.922	0.939	0.963	0.973
	MAE/(°C)	0.101	0.073	0.065	0.057	0.044

, and the model prediction results are illustrated in Figure 9d. The results indicate that Model 5 exhibits a lower prediction error for soil temperature, with an R² of 0.973 and an MAE of 0.043 °C for the prediction set. Compared to Models 1, 2, 3, and 4, Model 5 improves R² by 10.87%, 5.53%, 3.64%, and 1.02%, respectively, and reduces MAE by 56.72%, 40.33%, 32.62%, and 23.16%, respectively.

As illustrated in Figure 10, we have systematically compared the performance of five models in predicting four key factors within a greenhouse environment. By constructing a comprehensive performance evaluation chart encompassing MAE, MAPE, MSE, CV-RMSE, and R², we can visually assess the predictive capability of each model more intuitively.

The results unequivocally demonstrate that Model 5 significantly surpasses the other four comparative models across all pertinent performance indicators. Model 5 not only exhibits reduced prediction biases in MAE and MAPE but also demonstrates a narrower range of error variations in MSE and CV-RMSE evaluations, thereby indicating more stable and reliable prediction outcomes. Concurrently, Model 5’s exceptional performance in R² further validates the high linear correlation between its predicted values and the actual values, highlighting its superiority in prediction accuracy.

In addition to comparing prediction performance, we also evaluated the computational efficiency of each model. The average computation times for Models 1 to 5 are 52 s, 50 s, 77 s, 128.4 s, and 81.1 s, respectively. Although Model 5 excels in prediction accuracy, its computational efficiency is relatively slower.

From the overall prediction results, compared to Models 1 and 2, we observed that, compared to LSTM, BiGRU was able to better extract long-term dependencies in time series data, thus exhibiting superior performance in handling such data. Model 1, serving as the baseline model, demonstrates a relatively fast computation speed but lower prediction accuracy. In contrast, Model 2 achieves a slight improvement in prediction accuracy while maintaining a comparable computational speed. Therefore, it is a reasonable choice to incorporate BiGRU as a component of the prediction model. By introducing the CNN module before the BiGRU model in comparison to Models 2 and 3, we were able to extract features and enhance the spatiotemporal correlation of environmental data. The computation time for Model 3 increases to 77 s, yet its prediction accuracy is further enhanced. This observation suggests that feature extraction exerts a noticeable impact on computation speed. Furthermore, compared to Models 3 and 4, adding the SAM after the overall model can help the model better capture the key information in the sequence data, improve the model’s attention to different parts of the sequence data, and perform better in the time series data prediction task. However, the incorporation of SAM increases the model’s computation time to 124.8 s, making it the most computationally costly among all the models. Finally, compared to models 4 and 5, adding the GJO swarm intelligence optimization algorithm before the overall model enables the data to be searched and updated at different levels through collaborative optimization to better explore the global optimal solution in the solution space, effectively optimize the model parameters, obtain the best overall prediction, and improve the accuracy of the model prediction. Despite the relatively long computation time of 81.1 s for Model 5, its exceptional prediction accuracy and stability justify the design rationale, offering robust support for precise prediction of greenhouse environmental factors.

5. Discussion

The solar greenhouse constitutes a vital agricultural production instrument in northern China, crucial for achieving high-yield crop production, owing to its capability to maintain a controlled internal environment and its resilience amidst seasonal variations [42]. This study undertakes a comprehensive analysis of the disparities observed in the interior environmental parameters of greenhouses. By incorporating the findings from the GCBS-based prediction model, it subsequently elucidates the intricate coupling relationships among these parameters and explores their potential implications for the efficient management of greenhouses.

In the disparity analysis, we uncover significant coupling effects among the greenhouse interior environmental parameters. The interplay between parameters, including light intensity, carbon dioxide concentration, air temperature, relative humidity, and soil temperature, collectively contributes to maintaining the ecological balance within the greenhouse. This intricate coupling not only corroborates the direct impact of biological processes like photosynthesis on environmental parameters but also underscores the complexity of managing the greenhouse environment [43].

In pursuit of precise regulation of the greenhouse interior environment, our proposed GCBS prediction model has demonstrated remarkable advantages in forecasting greenhouse environmental parameters. Compared to other similar models, the GCBS model exhibits significant superiority in prediction accuracy, computational efficiency, and generalization capabilities. Specifically, the GCBS model adeptly captures the varying trends of highly dynamic carbon dioxide concentrations, exhibits heightened sensitivity in predicting subtle fluctuations in air relative humidity and consistently maintains high accuracy and stability in predicting periodic variations in air and soil temperatures. The robust data processing capabilities and highly precise prediction outcomes of the GCBS model provide robust technical support for the scientific management of greenhouse environments.

However, the GCBS model entails a certain time cost during the initial training phase. Given that the GJO optimization algorithm is an iterative search method, it necessitates multiple iterations to gradually converge toward optimal parameters, which can be time-consuming. Addressing these limitations, we plan to incorporate various optimization strategies in future research to further reduce training time costs while continuously enhancing model performance.

Throughout the research process, we have systematically gathered greenhouse climate and environmental data spanning from winter to early spring and have planned to further expand the dataset in future work by incorporating data from summer and autumn, as well as equipment operational data. This will assist in dividing the dataset by season and developing prediction models tailored to each season, enabling a more accurate capture of the characteristics of these seasonal variations and subsequently enhancing the effectiveness and practicality of the prediction models. This research outcome will provide reliable environmental prediction tools for greenhouse agricultural production, better assist agricultural production practices, and provide crucial support for optimizing greenhouse environmental management and increasing crop yields.

6. Conclusions

This study proposes a multi-parameter and multi-node prediction model, namely GCBS, for the prediction of environmental parameters in solar greenhouses. This model, grounded in a thorough analysis of the spatial-temporal distribution and coupling relationships of the primary environmental parameters within a greenhouse, selects eight crucial parameters as inputs: outdoor air temperature, outdoor air relative humidity, outdoor soil temperature, indoor light intensity, indoor air temperature, indoor air relative humidity, indoor CO₂ concentration, and indoor soil temperature. By leveraging these inputs, the model achieves accurate predictions for the upcoming hour about the air temperature and humidity, soil temperature, and CO₂ concentration. The experimental findings conclusively demonstrate the remarkable predictive accuracy of the GCBS model. For predictions of CO₂ concentrations, the GCBS model attained an R² value as high as 0.99, concurrently reducing the MAE to 19.50 mg/kg, diminishing the MAPE to 2.60%, lowering the MSE to 994.20 (mg/kg)², and reaching a low of 4.34% for CV-RMSE. Similarly, the GCBS model exhibited exceptional performance in predicting three additional environmental parameters, with R² values surpassing 0.96 for all parameters. Furthermore, key metrics such as MAE, MAPE, MSE, and CV-RMSE consistently outperformed those of alternative models, thereby substantiating the high stability and accuracy of the GCBS model. The model integrates the benefits of optimization algorithms and multi-level network structures, effectively capturing temporal relationships between data and aptly handling the complex coupling of environmental parameters, thereby providing a powerful tool for precise prediction of solar greenhouse environments. Although the GCBS model exhibits excellent performance in prediction accuracy, there is still room for further improvement in terms of optimization speed and computational efficiency. In the future, we plan to further optimize model parameters and network structures to adapt to prediction demands under varying environmental conditions, enhance the model’s practicality and generalization capabilities, and offer more comprehensive and efficient support for greenhouse environmental management and crop production.