The Intelligentization Process of Agricultural Greenhouse: A Review of Control Strategies and Modeling Techniques

Abstract

With the increasing demand for sustainable food production, the facility agriculture is progressively developing towards automation and intelligence. Traditional control techniques such as PID, fuzzy logic, and model predictive control have been widely applied in greenhouse planting for years. Existing greenhouse management systems still face challenges such as limited adaptability to fluctuating outdoor climates, and difficulties in maintaining both productivity and cost-effectiveness. Recently, with the development of greenhouse systems towards comprehensive environmental perception and intelligent decision-making, a large number of intelligent control and modeling technologies have provided new opportunities for the technological update of greenhouse management systems. This review systematically summarizes recent progress in greenhouse regulation and crop growth control technologies, emphasizing applications of intelligent techniques, involving adaptive strategies, neural networks, and reinforcement learning. Special attention is given to how these methods improve system robustness and control performance in terms of environmental stability, crop productivity, and energy efficiency, which are key performance indicators of greenhouse systems. Their advantages over conventional strategies in agricultural greenhouse systems are also analyzed in detail. Furthermore, the integration of intelligent technologies with greenhouse system modeling is examined, covering both greenhouse environmental models and crop growth models. The strengths and weaknesses of different techniques, such as mechanism, computational fluid dynamics (CFD), and data-driven models, are analyzed and discussed in terms of accuracy, computational cost, and applicability. Finally, future challenges and research opportunities are discussed, emphasizing the need for real-time adaptability, sustainability, and cluster intelligence.

1. Introduction

As a semi-closed ecosystem, a greenhouse allows for the regulation of various variables such as climate and nutrition that affect crop growth and development [1]. With the goal of producing the highest quality products at the lowest cost, the optimal planting conditions at main stages of crop growth could be achieved more precisely in greenhouses than field planting [2,3]. However, as the main element of a greenhouse, crops are influenced by various internal and external factors, such as external meteorological conditions (temperature, humidity, solar radiation, and carbon dioxide), nutrient levels (water and nutrients), and diseases (pests and viruses) [4,5]. There are multiple nonlinear coupling relationships between greenhouse crops and their various influencing elements, making an ecosystem with high complexity [6].

To this end, adequate control strategies are needed to keep the key factors (variables) within the optimal ranges for crops growth [7]. By now, it is well understood that greenhouse planting contains multiple simultaneous processes with different timescales and patterns. Numerous control systems for greenhouse regulation with different key performance indicators (KPIs) have been successfully applied [8,9,10,11]. According to various physical, chemical or biological processes within the greenhouse and their different time scales/KPIs, a multilayer hierarchical control architecture is generally constructed, which has three layers as follows [12,13].

(i) The upper layer mainly deals with long-term market goals, including market prices, overall returns, and energy costs.

(ii) The middle layer mainly deals with the growth process of crops throughout their entire life cycle. The controlled KPIs are physiological and production related indicators at slow timescale, such as leaf area index (LAI), canopy photosynthetic efficiency, dry matter accumulation, yield per unit area, and product quality.

(iii) The lower layer is mainly responsible for climate and irrigation control at each time step of crop growth, corresponding to fast timescale in seconds to minutes. Resource efficiency is emphasized simultaneously, including water and nutrient use efficiency, and the ratio between energy input and biomass production. The schematic diagram of hierarchical control architecture is shown in Figure 1.

This paper conducts a comprehensive review of recent studies on intelligent technologies in agricultural greenhouse systems, with emphasis on the control strategies and modeling techniques. The review scope mainly corresponds to the middle and lower layers of the schematic framework, while issues related to market pricing and economic returns are excluded.

Relevant publications are retrieved through systematic searches in Web of Science^®, Google Scholar^®, ScienceDirect^®, China National Knowledge Infrastructure (CNKI)^®, and SpringerLink^®. Search strings combine keywords such as “greenhouse environment” AND “intelligent control”, “artificial intelligence” OR “neural networks”, and “reinforcement learning”. The search covers the period from 2010 to 2025, with the last update conducted in March 2025. Inclusion criteria requires publications to be peer-reviewed journal articles or high-quality conference papers. Articles are primarily limited to English; however, a small number of Chinese publications from CNKI are included when they provide English abstracts and contain highly relevant technical content. Exclusion criteria includes: (i) studies focusing on open-field agriculture rather than greenhouses, (ii) papers lacking methodological or performance details, and (iii) reviews or editorials without original results. After duplicate removal, titles, abstracts, and full texts are screened stepwise. In total, 106 relevant research articles are retained for analysis, forming the basis of this review.

All reviewed articles cover control strategies for various greenhouse subsystems, such as greenhouse microclimate control, irrigation control, crop growth control, etc. From the perspective of automatic control, all reviewed articles can be classified according to controller principles and control variables.

(1) Controller principles In the lower layer of greenhouse system, most traditional controllers belong to set point control (such as switch controller or PID controller). In these control strategies, the ideal set value (objective) is given by expert experience or derived from upper level optimization algorithms. Intelligent technologies are mostly introduced to enhance the dynamic response and parameter settings of these lower layer controllers. For crop growth and entire greenhouse climate control in the middle layer, optimal control strategies are commonly applied, in which a comprehensive target taking into account crop growth and energy consumption is needed. For these controllers, the intelligent technologies are more necessary, especially for some simulation based sub models. Intelligent technologies may play an important role in global parameter optimization, iterative solving of optimal control sequences, and other aspects [14].

(2) Control variables For greenhouse environment system, the common controlled variables include air temperature, humidity, solar radiation, and carbon dioxide. For greenhouse irrigation system, water and nutrients are the main controlled variables. Besides, sunshade rate, fan speed, wet curtain temperature, etc., are also set as control variables in some sub-models in greenhouses. Beyond maintaining favorable microclimatic conditions, greenhouse systems are expected to achieve high energy efficiency, ensure crop yield/quality, and demonstrate adaptability to external climate variations. These KPIs provide explicit and measurable objectives for evaluating intelligent control and modeling approaches.

Out of the 106 selected articles, 33 papers addressed conventional control strategies, mainly involving three types of controllers: PID control, fuzzy control, and model predictive control (MPC). A total of 34 studies focused on intelligent control strategies, among which 17 applied neural network–based control (NNs control), 5 adopted reinforcement learning–based (RL-based) control, and 12 employed other intelligent control techniques such as adaptive control, feedback linearization, and event-based control. It should be noted that for articles involving multiple methods, only the primary or emphasized method is counted in order to avoid double-counting. The classification is carried out manually by two reviewers, with any disagreements resolved through discussion until consensus is reached. Figure 2 and Figure 3 present the distribution of literature on conventional and intelligent control strategies, respectively, in the form of bar charts to more appropriately reflect the number of studies in each category.

This review aims to discuss recent research on intelligent technology in agricultural greenhouse environment control. In the past few years, about seven review papers on intelligent greenhouse systems have been published. For example, several works have surveyed conventional control strategies such as PID and fuzzy logic, while others focused on specific modeling approaches including mechanistic and CFD simulations. However, a comprehensive review that systematically integrates advanced intelligent technologies across modeling, control, and crop growth aspects remains lacking. To clarify the research gaps in existing reviews,

Table 1. Summary of previous review papers, since 2022.

Contents	[1] (2022)	[2] (2022)	[3] (2023)	[4] (2024)	[5] (2024)	[6] (2025)	[7] (2025)
Conventional control techniques					×
Neural Network based Control
Reinforcement learning	×	×	×	×		×	×
Data-driven modeling					×
Intelligent Technologies in Greenhouse Environment Modeling		×	×	×	×
Intelligent Technologies in Crop Growth Modeling		×	×	×	×	×	×

is provided, which compares previous works based on the key themes of this review: (I) conventional control techniques; (II) intelligent control methods; (III) mechanistic and data-driven modeling; (IV) CFD-based environmental simulation; (V) crop growth models. From Table 1, it is observed that most previous reviews did not sufficiently explore advanced intelligent control methods such as RL and data-driven modeling. Moreover, few studies have provided an integrated overview of both environmental modeling and crop growth modeling within the same intelligent technology–driven framework. In this paper, we address these underexplored areas and present a systematic analysis of intelligent control technologies applied to greenhouse environmental regulation and modeling. The main contributions of this review are: (1) a structured overview of both conventional and intelligent control strategies; (2) a detailed examination of intelligent technology–based modeling approaches for greenhouse environments, including mechanistic, CFD, and data-driven models; and (3) a comprehensive review of crop growth models in controlled environments. Furthermore, the challenges and future research directions are discussed.

2. Control Techniques in Agricultural Greenhouse Systems

Within modern agricultural greenhouse systems, various variables are particularly critical to crop growth, including air temperature, relative humidity, $C{O}_2$ concentration, and light intensity. In addition, water and nutrient supply in irrigation system, as well as ventilation rate, shading degree, and heating or cooling capacity, also need precise regulation. In this section, we review and discuss relevant articles based on the types of control principle, which could be mainly divided into two categories: conventional control and intelligent control. Under the same type of control principle, articles will be classified according to different controlled variables.

2.1. Conventional Control Techniques

In the past decades, a variety of conventional control techniques were successfully applied in greenhouse system control, such as PID control, fuzzy logic control, and MPC. Generally, the input-output relationships of the greenhouse system could be summarized in a brief schematic diagram as shown in Figure 4. This section will emphasize conventional controllers’ operating principles, representative applications, and limitations, which motivate the subsequent integration of intelligent technologies.

2.1.1. PID Control

As a classical and widely applied technique in greenhouse control, PID method can be traced back to the theoretical framework first proposed by Nicolas Minorsky in 1922 [15]. The fundamental mathematical model is established through three basic control actions: proportional (P), integral (I), and derivative (D). The specific formulation is given as follows:

$$u\left(t\right)=K_{p}e\left(t\right)+K_i{\int }_0^{t}e\left(\tau \right)d\tau +K_d{\displaystyle \frac{de\left(t\right)}{dt}}$$

(1)

where $e\left(t\right)$ denotes the error between the set point and actual value, and $K_p$, $K_i$, $K_d$ represent the proportional, integral, and derivative coefficients, respectively. The relevant principle is also illustrated in Figure 5.

Due to the simplicity and ease of implementation, PID has been widely used for various variables’ regulation in agricultural greenhouse systems [16]. Considering the compactness of content, we only select representative PID control articles from the past 10 years, and review them by their main controlled variables as follows.

(1) Temperature/humidity control As the dominant variables for crop growth in greenhouse environment, temperature/humidity regulation is the most extensive applications for PID control. For instance, Gao et al. [17] proposed a coupled temperature/humidity PID controller that incorporated cross-variable compensation, reducing stabilization times by 73.3% for temperature and 50% for humidity. Bao et al. [18] designed a PID based temperature/humidity controller for grape growth, and achieved steady accuracy of 80–90%. It is noted that conventional PID controllers with fixed parameters often fail to handle nonlinear dynamics, time delays, and sudden disturbances such as rapid weather changes. Hu et al. [19] employed the evolutionary algorithm to tune PID controller’s parameters. By deriving a Pareto-optimal solution set, their method achieved a balance between temperature–humidity control performance and actuator wear. Su et al. [20] introduced a parameter self-tuning PID method. The controlled variables of temperature, humidity, and $C{O}_2$ were transformed into equivalent single variable for dynamic decoupling while maintaining accuracy and reducing actuator switching frequency. Compared with temperature control, humidity’s PID regulation remains challenging because of its delayed response and the mutual influence with crop physiological processes.

Note: The reported improvements are drawn directly from the cited studies, where experimental conditions and baselines were defined, and thus their applicability is bounded by the respective simulation or field settings. In fact, all comparative statements in this review follow the same principle, i.e., they are based on the original studies’ defined benchmarks and contexts rather than unconditional superiority claims.

(2) $C{O}_2$ concentration control The $C{O}_2$ concentration directly influences photosynthesis and crop yield in greenhouses, thus its precise regulation is also critical. Berenguel et al. [21] combined feedback linearization with PI controller and used a genetic algorithm for parameter tuning, achieving stable $C{O}_2$ regulation even under parameter disturbances up to $\pm 10$%. Su et al. [20] further enhanced PID-based control through adaptive mechanisms, improving $C{O}_2$ concentration accuracy by 35% while saving 23% energy. Despite these advances, spatial heterogeneity of $C{O}_2$ distribution inside greenhouses requires more adaptive and distributed-parameter solutions.

(3) Irrigation and nutrient control In water–fertilizer management, PID controllers have been widely used for regulating pump speeds and valve operations [22,23,24,25,26,27]. Li et al. [28] developed a fuzzy adaptive PID algorithm with iterative learning compensation for fertigation systems, enabling real-time adjustment and pre-compensation of dosing deviations. Yuan et al. [29] designed a PID-based fertigation control system for greenhouse cucumbers, which integrated soil moisture and nutrient sensors to regulate irrigation valves and fertilizer dosing pumps. Their results showed that the system maintained soil moisture within a $\pm 2$% error range and nutrient concentration deviations within $\pm 5$%, effectively improving water–fertilizer use efficiency while reducing manual intervention.

(4) Light intensity and shading control Although less reported compared with above variables, PID controllers were also applied in adjusting shading devices and supplemental lighting systems to regulate light intensity required by crop growth [30]. In practice, light regulation is often coupled with temperature and $C{O}_2$ control to optimize photosynthesis [31,32,33]. However, rapid fluctuations in natural solar radiation present difficulties for fixed-parameter PID controllers, making hybrid or intelligent methods more attractive.

2.1.2. Fuzzy Control

In the field of agricultural greenhouse system control, fuzzy control represents one of the classical conventional control techniques, with its theoretical foundation traced back to the fuzzy set theory proposed by Zadeh in 1965 [34]. It simulates human fuzzy reasoning and decision-making processes: converts precise inputs into fuzzy linguistic variables, conducts reasoning according to preset rules, and finally outputs precise control quantities to achieve effective control of complex and nonlinear systems. Three main parts are involved: fuzzification, fuzzy inference and defuzzification. A detailed principle introduction of typical fuzzy logic controller could be found in [35], and a basic schematic diagram is shown in Figure 6.

Unlike conventional PID control, fuzzy control defines membership functions for various environmental variables, and incorporates expert knowledge through “IF–THEN” fuzzy rules to naturally express and coordinate complex coupling relationships in greenhouse systems. Recent studies enhanced the adaptability of fuzzy control under time-varying environments by introducing wireless sensor networks and online learning mechanisms [36]. Azaza et al. [37] proposed a dynamic adjustment method for membership functions based on real-time monitoring data. The method achieved 22% energy savings and a 33% increase in water-use efficiency during the coordinated regulation of multiple variables (temperature, humidity, light, etc.), significantly outperforming conventional PID control in similar complex scenarios. Wang et al. [38] further incorporated the accumulated temperature theory to design a variable-universe fuzzy controller. By dynamically adjusting the fuzzy domain range via scaling factors, their approach reduced overshoot by 75% during the seedling-to-flowering stage, while energy consumption decreased by 10% compared with conventional fuzzy control.

Beyond structural optimization and parameter self-adaptation, intelligent learning algorithms have also been applied to enhance fuzzy controller’s performance. El Aoud et al. [39] proposed a fuzzy adaptive controller based on a gradient descent algorithm, which dynamically adjusted membership functions and rule weights through online learning. This method effectively reduced controller complexity (in terms of rule quantity and parameter adjustment burden) and achieved satisfactory trajectory tracking performance in greenhouse temperature/humidity regulation. Su et al. [40] developed a fuzzy adaptive composite control system that integrated feedback linearization with Lyapunov stability analysis. Their scheme improved $C{O}_2$ concentration control accuracy by 35% and achieved 23% energy savings by relaxing the tracking error threshold. Xing et al. [41] constructed a dual fuzzy controller architecture, employing a ZigBee network for distributed acquisition of environmental parameters. Through day/night segmented optimization, they achieved a temperature control accuracy of ±0.5 °C, significantly surpassing the ±1.2 °C accuracy of conventional PID control. Wei et al. [42] designed a three-input fuzzy expert system that incorporates second-order variations of environmental parameters to establish a multi-dimensional reasoning mechanism. In tomato growth regulation, their system improved response speed by 40% and steady-state accuracy by 25%.

2.1.3. Model Predictive Control

Unlike PID control and fuzzy control, MPC leverages a multi-variable cooperative optimization mechanism grounded on dynamic models, thereby overcoming the limitations of traditional single-loop regulation. Its theoretical foundation can be traced back to Dynamic Matrix Control (DMC), proposed by J. Richalet and his colleagues in 1978 [43], and Model Algorithmic Control (MAC), developed by C. R. Cutler in 1982 [44]. MPC computes optimal control actions by predicting the future behavior of a system over a finite horizon using a dynamic model. It solves an online optimization problem at each time step and applies only the first control input, then repeats the process with new measurements to handle uncertainties. The basic principles of MPC comprise three key elements: dynamic prediction, receding (rolling) optimization, and feedback correction.

In the field of greenhouse environmental control, MPC achieves precise regulation of crop growth conditions by coordinating multi-variable constraints such as temperature, humidity, light, and $C{O}_2$ concentration [45,46]. For example, Chen et al. [47] proposed a MPC strategy for greenhouse temperature management, which effectively addressed uncertainties and disturbances in nonlinear systems to achieve high-precision tracking and enhanced robustness. Qi et al. [48] developed a TRNSYS-MATLAB co-simulation platform, enabling conducting MPC strategy by dynamically adjusting greenhouse temperature reference values. Results showed that the optimized MPC improved the average greenhouse temperature by 10.6% compared with traditional PID control during cold months, and reduced constraint violation rates by 29.7% through rolling horizon optimization. How to establish accurate dynamic models for greenhouse environment regulation is the key issue for MPC. Numerical simulation and data-driven modeling are the core approaches [49,50,51]. Mahmood et al. [52] combined analytical energy-balance models with artificial neural networks (ANN) to build a dynamic temperature predictive model of greenhouse in arid regions, based on which a robust MPC framework is implemented for temperature control. Their data-driven predictor achieved superior accuracy, with a root mean square error (RMSE) of 0.285–0.314 °C, compared with mechanistic models that exhibited an RMSE of 0.701–0.941 °C, thereby enabling precise temperature tracking under sandstorm-induced solar irradiance fluctuations. Further, They proposed an improved data-driven model predictive control strategy, which employed a dual-layer structure, in which the primary controller generated the nominal trajectory and the auxiliary controller dynamically compensated for uncertainties to optimize greenhouse temperature regulation [53]. Compared with conventional MPC, this approach substantially improved accuracy, yielding a mean absolute error (MAE) of 0.09 °C in winter, and reduced energy consumption with savings of 13.34% in summer. The schematic diagram of the robust model predictive control (RMPC) strategy is shown in Figure 7.

The summary of reviewed conventional control techniques is provided in

Table 2. Summary of conventional control techniques.

Reference	Type of Control	Control Variables	Study Setting	Control Performance
Gao et al. [17]	PID	Air temperature/Humidity	Lab	Stabilization time reduced by 73% (T), 50% (H)
Bao et al. [18]	PID	Air temperature/Humidity	Field	80–90% steady accuracy
Su et al. [20]	Self-tuning PID	Air temperature/Humidity	Lab	Maintained accuracy, fewer actuator switches
Berenguel et al. [21]	PI + feedback linearization	$C{O}_2$	Sim. + Lab	Stable under ±10% disturbances
Su et al. [20]	Adaptive PID	$C{O}_2$	Lab	Accuracy +35%, energy −23%
Zhang et al. [22];
Fu et al. [23]; Zhu et al. [24]	PID	Water/Nutrients	Field	Widely applied in fertigation
Li et al. [28]	Fuzzy adaptive PID	Water/Nutrients	Sim.	Real-time dosing adjustment
Yuan et al. [29]	PID	Water/Nutrients	Field	Soil error ±2%, nutrient error ±5%
Guan et al. [30]	PID	Light/Shading rate	Field	Applied in shading/lighting
Azaza et al. [37]	Fuzzy (dynamic)	Multiple variables	Sim.	Energy −22%, water-use +33%
Wang et al. [38]	Variable-universe fuzzy	Air temperature	Lab	Overshoot −75%, energy −10%
El Aoud et al. [39]	Fuzzy adaptive (GD)	Air temperature/Humidity	Sim.	Reduced rule complexity, better tracking
Su et al. [40]	Fuzzy adaptive composite	$C{O}_2$	Lab	Accuracy +35%, energy −23%
Xing et al. [41]	Dual fuzzy	Air temperature	Lab	Accuracy ±0.5 °C vs. ±1.2 °C (PID)
Chen et al. [47]	MPC	Air temperature	Sim. + Lab	High-precision tracking, robust
Qi et al. [48]	MPC	Air temperature	Sim. + Field	Accuracy +10.6%, violations −29.7%
Mahmood et al. [52]	Robust MPC	Air temperature	Sim.	RMSE 0.29–0.31 °C
Mahmood et al. [53]	Dual-layer data-driven MPC	Air temperature	Sim. + Field	MAE 0.09 °C (winter), energy −13.3% (summer)

, in which the control methods, objects, and performances are categorized.

2.2. Intelligent Control Technology

Although the aforementioned conventional control techniques (e.g., PID, fuzzy control, and MPC) have achieved remarkable success in specific scenarios, their control performance remains highly dependent on accurate mechanistic models or expert knowledge [54,55,56,57]. With the rapid development of artificial intelligence technologies, intelligent control methods have emerged as promising alternatives to address these challenges [58,59,60]. This section systematically reviews the applications of intelligent control techniques in agricultural greenhouse systems. The focus is put on how these methods enhance system robustness and control performance, while also analyzing their advantages over conventional control strategies.

2.2.1. Neural Networks-Based Control

Inspired by the connectivity and learning mechanisms of biological neurons, neural networks provide an effective approach for addressing the nonlinearity, time-variability, and multivariable coupling challenges of greenhouse systems [61,62]. Depending on network architecture and temporal modeling capacity, NNs can be divided into feedforward, recurrent, deep temporal networks, etc.

(1)

Back Propagation (BP) neural networks

BP neural networks are widely used for modeling and optimization tasks due to its nonlinear mapping capacity and simple structure [63]. It has a multi-layer feed forward structure based on error backpropagation algorithm, which optimizes weights through gradient descent to achieve nonlinear mapping from input to output. A basic three-layer BP structure is described in Figure 8.

In applications of greenhouse system control, BP networks are frequently used to capture complex couplings among variables such as temperature, humidity, light intensity, and $C{O}_2$ concentration. Compared with conventional control manners (such as PID control), BP networks offer superior accuracy and robustness in prediction and adaptive regulation, particularly under sensor noise and external climatic disturbances. Castaneda et al. [64] developed a BP networks as state estimator for frost control in greenhouses. The designed three-layer network structure achieved precise temperature prediction with 95% confidence, reducing errors by over 50% compared with a classical ARX model. Differently, Dingguo [65] applied BP networks for tuning of PID coefficients. Instead of gradient descent, the particle swarm optimization (PSO) was used to optimize weights of the BP networks. The obtained BP-PSO-PID controller reduced temperature adjustment time by 82.3% and overshoot by 67% compared with BP-PID. Similarly, Jianping et al. [66] employed a GA-PSO-BP hybrid model, cutting overshoot and adjustment time by 35% and 40% in simulations. For specific crops’ variable control, Feng et al. [67] applied an optimized BP-PID controller in tomato greenhouses, maintaining stable temperatures (22.5–26.1 °C) and reducing humidity fluctuations by 14.3%. Xinxin et al. [68] designed heterogeneous networks for Oncidium (3-12-3 structure) and Phalaenopsis (3-10-3 structure), achieving prediction RMSEs of temperature, humidity, and light intensity by 1.4 °C, 5.1%, and 3.0 klx, respectively. Liqun [69] further improved adaptability by introducing a B-BP network with $\beta$-parameterized B-spline basis functions, outperforming BP-PID and RBF-PID in both overshoot and response speed.

(2)

Recurrent neural networks (RNN)

Unlike feedforward neural networks, RNN introduce “recurrent” connections in the network, which means when processing the current input, the network takes into account the historical information of all previous inputs. Due to this characteristic, it is very suitable for time series data’s prediction and control. The basic principle of RNN is shown in Figure 9.

Elman neural network is one of the simplest recurrent neural network, which introduces a context layer to form a recurrent architecture with temporal memory. Aytenfsu et al. [70] applied the Elman neural network to predict inner temperature and humidity of a greenhouse by using four input variables (wind speed, $C{O}_2$, outdoor temperature, and humidity). Similarly, Zhang et al. [71] employed Elman networks to achieve more accurate predictions of temperature, humidity, and $C{O}_2$ compared with BP and RBF networks, with a mean squared error (MSE) of 0.0039 and a coefficient of determination ($R^2$) of 0.9915 for temperature. For controller design, Belhaj Salah et al. [72] trained an Elman neural network to emulate the thermal dynamics of greenhouse system. It was cascaded with a deep multi-layer perceptron (MLP) as controller of greenhouse internal climate. Similar study could be found in [73]. Fourati et al. [74] developed a forward dynamic model using an Elman neural network in combination with a feedforward neural network (FFNN) inverse controller, achieving substantial reductions in temperature and humidity tracking errors and highlighting the inherent advantages of neural networks in managing time-varying systems and external disturbances.

To overcome parameter sensitivity and local optima, various optimization algorithms were integrated with Elman networks. For example, Pan et al. [75] optimized Elman networks using a sparrow search algorithm (SSA). On winter solar greenhouse datasets from Shandong, China, the SSA-Elman achieved RMSE values of 0.592 ($R^2$ = 0.963) for temperature and 2.530 ($R^2$ = 0.972) for humidity, improving prediction by over 7% than the baseline Elman.

(3)

Long short-term memory (LSTM) neural networks

LSTM networks extend conventional RNNs by introducing gating mechanisms (i.e., Forget, Input, and Output gates) to regulate information flow through the cell state, which alleviates the vanishing gradient problem and enables robust long-sequence learning [76]. Compared with standard RNNs, the memory-preserving mechanism of LSTM provides an engineering-feasible solution for greenhouse control tasks requiring long-horizon dependency modeling. The basic LSTM unit with gating mechanisms is shown in Figure 10. In the schematic diagram, $C_t$ and $h_t$ indicate the cell state and hidden state at current time t, and $\sigma$ represents activation function of fully connected layer. More details could be found in related reference such as [77].

For multi-variable prediction of greenhouse environmental systems, Chen et al. [78] demonstrated that a two-layer LSTM with a 15-min input window achieved notable improvements in temperature, humidity, and $C{O}_2$ prediction compared with RNNs and gated recurrent units (GRUs). Ali et al. [79] developed a two-layer LSTM-RNN for greenhouse climate prediction, trained on seven years of German temperature data (N = 420,551). By optimizing hidden units (200), network depth (2 layers), and activation functions, the model achieved an RMSE of 0.069 in 12-h prediction tasks. Jung et al. [80] analyzed 470,000 multisource data records from a Korean greenhouse, including environmental and actuator variables. They demonstrated the superiority of RNN-LSTM over ANN and NARX models for long-horizon prediction. Their ablation experiments further quantified the importance of actuator control history for humidity prediction ($R^2$ dropped by 0.03 upon removal), showing an influence three times stronger than external meteorological inputs. Recently, Gong et al. [81] proposed a hierarchical temporal feature extraction framework that integrated LSTM-RNN modules for sequence learning with temporal convolutional network (TCN) dilated convolutions for multi-scale feature fusion. Using cross-annual UK greenhouse datasets (2017–2018), the hybrid model achieved up to a 29.5% reduction in RMSE compared with multi-layer LSTM-RNNs. To enhance short-term responsiveness, Qiao et al. [82] proposed an LSTM-GRU hybrid with Kalman filtering, achieving 93.9% of temperature predictions within ±0.5 °C and reducing RMSE and MAE by over 20%. To further enhance transparency and comparability, we have provided additional contextual information for neural network-based control studies. We summarize the study settings and validation manners of the NN-based control studies in

Table 3. Study settings and validation protocols of NNs-based control studies.

Reference	Study Setting	Dataset Size/Duration	Train/Test Split
BP Neural Networks
Castaneda et al. [64]	Field	2880 samples (10-min)/1 year (summer & winter)	70/30 split
Dingguo [65]	Sim.	600 samples (Not reported interval)/Not reported duration	Not reported
Jianping et al. [66]	Sim.	Not reported	Not reported
Feng et al. [67]	Sim. + Field	Not reported/1 day	70/30 split
Xinxin et al. [68]	Field	438 samples (10-min)/multi-season (Oncidium); 371 samples (10-min)/multi-season (Phalaenopsis)	78/22 split
Liqun [69]	Sim.	600 samples (0.01 s)/Not reported duration	Not reported
Recurrent Neural Networks (RNN)
Aytenfsu et al. [70]	Field	18,000 samples (5-min)/10 days	75/25 split
Zhang et al. [71]	Field	2880 samples (3-min)/5 days	2400/480 split
Belhaj Salah et al. [72]	Sim.	1440 samples (1-min)/1 day	50/50 split
Fourati et al. [74]	Field	1440 samples (1-min)/1 day	Not reported
Pan et al. [75]	Field	3500 samples (30-min)/93 days	80/20 split
Long Short-Term Memory (LSTM)
Chen et al. [78]	Field	2880 samples (10-min)/1 year	70/30 split
Ali et al. [79]	Sim.	600 samples (Not reported interval)/Not reported duration	Not reported
Jung et al. [80]	Sim. + Field	Not reported/1 day (12 h)	70/30 split
Gong et al. [81]	Field	Not reported/1 year	70/30 split
Qiao et al. [82]	Sim.	600 samples (0.01s)/Not reported duration	Not reported

. It details whether the studies were conducted under simulation, laboratory, or field conditions. The dataset size and division manner are also provided in this table.

2.2.2. Adaptive Control

Adaptive control strategies dynamically adjust controller parameters or structures online to accommodate the varying dynamics and external disturbances of greenhouse systems, thereby significantly enhancing control robustness. The essence of such methods lies in developing mechanisms capable of perceiving system states in real time and automatically correcting control laws. In recent years, intelligent control systems based on adaptive genetic algorithm improvements have demonstrated excellent performance in greenhouse environmental regulation [83]. For greenhouse environmental control, adaptive control is often integrated with classical PID algorithms or intelligent control methods to form high-performance hybrid control architectures. For instance, Li et al. [84] proposed a fuzzy adaptive PID algorithm that embeds expert knowledge into a fuzzy rule base to dynamically tune PID parameters, thereby enabling real-time and precise regulation of water-fertilizer mixing processes. Similarly, Zeng et al. [85] employed an RBF neural network to identify the system Jacobian online and dynamically adjust the gain parameters of a nonlinear adaptive PID controller, effectively mitigating multivariable coupling and nonlinearities in greenhouses and thereby improving real-time performance and stability. In recent years, the integration of hierarchical control architectures with data-driven approaches has offered new insights into addressing system uncertainties. Mansour et al. [86] proposed an adaptive robust control framework that combines upper-layer economic MPC with lower-layer deep reinforcement learning (DRL) agents. Within this framework, the upper-layer MPC generates economically optimal climate reference trajectories based on weather forecasts and energy prices, while the lower-layer DRL agent, trained through a two-stage learning strategy, acquires robust tracking capabilities and maintains stable operation under conditions of actuator faults and environmental disturbances. This approach markedly improves the adaptability and fault tolerance of greenhouse control systems under real-world operating conditions.

2.2.3. Feedback Linearization Control

Feedback linearization is a representative nonlinear control method whose core idea lies in transforming the original nonlinear system into an equivalent linear system through nonlinear state transformations and feedback compensation. This enables the direct application of well-established linear control theory to controller design. The method is typically suitable for nonlinear systems with well-defined dynamics that can be modeled with high accuracy [87,88]. In agricultural greenhouse environmental systems, variables such as temperature and humidity generally exhibit significant nonlinearities and strong couplings. For example, temperature and humidity mutually influence each other through evaporation and ventilation processes. This complexity provides a typical application scenario for feedback linearization. Chen et al. [89] further proposed a feedback linearization–based predictive control framework integrated with an Unscented Kalman Filter. This approach enabled online estimation of time-varying parameters for the nonlinear greenhouse model (e.g., the solar radiation conversion factor), and transformed a first-order physical model into an affine nonlinear system. When combined with MPC, the framework achieved ±1.0 °C tracking accuracy while optimizing energy consumption, thereby significantly enhancing adaptability.

Nevertheless, the practical application of feedback linearization in greenhouse environments still faces challenges, primarily due to the difficulty of obtaining precise and generalizable nonlinear analytical models. Current research trends are shifting toward hybrid approaches that integrate feedback linearization concepts with data-driven methods—for example, using neural networks to approximate the nonlinear components of the system, thereby achieving “soft feedback linearization”.

2.2.4. Event-Based Control

Event-driven sampling and control are also attracted attention due to their ability to significantly reduce information exchange among sensors, controllers, and actuators [90]. This approach could extend the lifetime of battery-powered sensors, decrease computational load of embedded devices, and alleviate network bandwidth demands. Unlike conventional time-triggered control schemes with fixed sampling periods, event-based control determines the triggering instants of control actions based on the dynamic evolution of system states.

In greenhouse climate control, Ferre et al. [91] integrated wireless sensor networks with event-driven sampling and control strategies to improve regulation efficiency. The triggering mechanisms were typically based on level-crossing rules or deadband conditions, whereby sensing and actuation were executed only when variable deviations exceeded prescribed thresholds, effectively optimizing resource usage. On the basis, Pawlowski et al. [92,93] introduced event-based Generalized Predictive Control (GPC), where the model-based predictive optimization was asynchronously recalculated in response to detect events rather than periodic cycles, thus embedding intelligent decision-making into the control loop. Subsequent refinements incorporated sensor and actuator side deadbands directly into the predictive control formulation, modeled as constraints within a hybrid optimization framework, which enabled a systematic tradeoff between control accuracy, communication load, and actuator wear [94]. Experiments demonstrated that the event-driven predictive approach not only preserved satisfactory climate regulation performance but also reduced control signal updates by more than 80% compared with conventional time-triggered schemes, yielding significant energy savings and extended actuator lifetime.

2.2.5. RL-Based Control

RL-based control learns optimal control policies through continuous interaction between Agent and Environment in a trial-and-error manner, with the core objective of maximizing long-term cumulative Reward. The Agent performs Actions and receives Rewards and new States as feedback from the Environment. This approach does not rely on precise mathematical models of the system, making it particularly suitable for complex dynamic systems that are highly nonlinear, strongly coupled, and difficult to model accurately, such as greenhouse climate control. A basic RL schematic is shown in Figure 11.

In greenhouse control, RL could directly optimize control action sequences, enabling effective coordination of multiple variables including temperature, humidity, light intensity, and $C{O}_2$ concentration, while simultaneously balancing crop growth benefits and energy consumption. In recent years, owing to its powerful function approximation capacity, RL has emerged as a research frontier in intelligent greenhouse control. Wang et al. [95] adopted the Deep Deterministic Policy Gradient (DDPG) algorithm to simultaneously optimize temperature, $C{O}_2$ concentration, and light intensity setpoints, significantly improving cumulative fresh weight yield of cucumbers in simulation environments. Zhang et al. [96] proposed a robust model-based RL framework that integrated an environment model as a virtual simulator to accelerate training, while introducing a sample dropout mechanism to optimize worst-case performance, achieving a 57% reduction in energy consumption and a 26.8% improvement in setpoint tracking accuracy in tomato greenhouse control. To enhance robustness under uncertain environments, Ajagekar et al. [97] combined robust optimization (RO) with Deep RL and proposed the RO-DRL framework, which demonstrated outstanding performance in handling weather uncertainties. Ban et al. [98] employed an Actor-Critic framework to directly perform adaptive control on nonlinear, black-box greenhouse systems, achieving more than 20-fold improvement in control stability under extreme disturbances compared with conventional PID control. Moreover, RL is often employed to optimize the parameters of traditional controllers. Adesanya et al. [99] utilized a Deep Q-Network to online-tune PID controller parameters for greenhouse HVAC systems, achieving multi-objective optimization of energy consumption and temperature regulation performance.

As a summary, all reviewed neural network control methods are listed in

Table 4. Summary of neural network-based control.

Reference	Method	Control Variables	Performance
BP Neural Networks
Castaneda et al. [64]	BP	Air temperature	Error reduced >50% vs. ARX
Dingguo [65]	BP-PSO-PID	Air temperature	Faster response, less overshoot
Jianping et al. [66]	GA-PSO-BP	Air temperature	Shorter adjustment time
Feng et al. [67]	BP-PID	Air temperature/Humidity	Stable T/H, lower fluctuation
Xinxin et al. [68]	BP	Air temperature/Humidity	RMSE ±1.4 °C, 5% H
Liqun [69]	B-BP	Air temperature/Humidity/$C{O}_2$	Better overshoot, faster response
RNN
Aytenfsu et al. [70]	Elman NN	Air temperature/Humidity	Accurate short-term prediction
Zhang et al. [71]	Elman NN	Air temperature/Humidity/$C{O}_2$	$R^2$ = 0.99 for T
Belhaj Salah et al. [72]	Elman+MLP	Air temperature/Humidity/$C{O}_2$	Improved stability
Fourati et al. [74]	Elman+FFNN	Air temperature/Humidity	Better tracking
Pan et al. [75]	SSA-Elman	Air temperature/Humidity	RMSE 0.59 (T), 2.53 (H)
LSTM
Chen et al. [78]	LSTM	Air temperature/Humidity/$C{O}_2$	Outperformed RNN/GRU
Ali et al. [79]	LSTM-RNN	Air temperature	RMSE = 0.069 (12 h)
Jung et al. [80]	LSTM	Multiple variables	Stronger long-horizon accuracy
Gong et al. [81]	LSTM+TCN	Multiple variables	RMSE ≈ 30% lower
Qiao et al. [82]	LSTM-GRU+Kalman	Air temperature	94% within ±0.5 °C
Reinforcement Learning
Wang et al. [95]	DDPG	Multiple variables	Higher cucumber yield
Zhang et al. [96]	Model-based RL	Air temperature/Humidity	57% energy saving
Ajagekar et al. [97]	RO-DRL	Air temperature/Humidity/$C{O}_2$	Robust, energy efficient
Ban et al. [98]	Actor-Critic	Multiple variables	>20× stability gain
Adesanya et al. [99]	DQN-PID	Air temperature/Humidity	Optimized PID, energy saving

, and other intelligent control techniques are listed in

Table 5. Summary of other intelligent control techniques.

Reference	Method	Control Variables	Study Setting	Performance
Adaptive Control
Chen [83]	GA-based adaptive	Air temperature/Humidity/$C{O}_2$	Sim.	Improved robustness under uncertainties
Li et al. [84]	Fuzzy adaptive PID	Water/Nutrients	Field	Real-time tuning, precise regulation
Zeng et al. [85]	RBF NN adaptive PID	Air temperature/Humidity	Sim.	Online Jacobian estimation, higher stability
Mansour et al. [86]	Hierarchical adaptive (MPC+DRL)	Multiple variables	Sim.	Robust to faults/weather, improved adaptability
Feedback Linearization
Gurban et al. [87]; Zengshuai et al. [88]	Feedback linearization	Multiple variables	Sim.	Enabled linear design for nonlinear dynamics
Chen et al. [89]	FL + UKF + MPC	Air temperature	Sim.	±1.0 °C tracking accuracy, optimized energy
Event-Based Control
Ferre et al. [91]; Pawlowski et al. [92,93,94]	Event-driven WSN sampling	Air temperature/Humidity/$C{O}_2$	Field/Sim.	>80% fewer updates, energy saving, longer actuator life

. It should be emphasized that the apparent superiority of one control method over another is often conditional on the application context. Neural-network–based approaches and reinforcement learning achieve superior adaptability and prediction accuracy, yet they demand large-scale datasets and extensive training, and are mostly validated through simulations or small-scale experiments. Conventional PID and fuzzy controllers remain advantageous in practice for low-cost deployment and scenarios with limited data availability. Therefore, the comparative statements in this paper should be understood within these domains of applicability rather than as unconditional superiority.

3. Intelligent Techniques in Greenhouse System Modeling

3.1. Greenhouse Environmental Model

As the core component of greenhouse system, inner environmental modeling plays a key role in greenhouse crop growth management and control [100]. According to the basic principles of modeling, it can be divided into three categories: mechanistic models, CFD simulations, and data-driven models.

3.1.1. Mechanistic Model

Early mechanistic models primarily constructed static equations based on thermodynamics and heat transfer principles [101]. Henten [102] developed a multiple-variable dynamic environmental model (in low-order differential form), incorporating crop photosynthesis and respiration as biological feedback mechanisms, for optimal control of lettuce growth. Subsequently, other dynamic greenhouse environmental models were developed, including the Vanthoor [103], Ooteghem [104], and KASPRO [105] models.

To address parameter calibration challenges, evolutionary optimization approaches have emerged as promising solutions. Herrero et al. [106] applied multi-objective evolutionary algorithms (e.g., NSGA-II) to simultaneously optimize model accuracy and complexity, while Guzmán-Cruz [107] used genetic algorithms to adaptively recalibrate heat transfer coefficients affected by material aging, significantly improving long-term robustness. Such intelligent optimization techniques alleviate the rigidity of conventional parameterized models, though at the expense of high computational cost. Their effectiveness remains contingent on the structural soundness of the underlying model. These models are most effective when: (1) high spatial resolution is not critical; (2) long-term prediction horizons are required; (3) prior physical knowledge is comprehensive; and (4) computational resources permit detailed parameter optimization. However, they are less adaptive to dynamic changes and require extensive calibration, which limits their practicality in large-scale or rapidly changing production scenarios.

3.1.2. CFD Simulation

To address the spatial distribution limitations of lumped parameter models, CFD has been widely used for greenhouse environmental modeling and optimization [108,109,110]. Such fine-grained numerical simulations based on physical fields provide high spatiotemporal resolution predictions of temperature and velocity fields, delivering critical environmental response relationships for dynamic system modeling and optimization [49].

Ni et al. [111] employed CFD simulations using ANSYS Fluent to analyze the sucrose flow field within tomato stems. The results revealed a pressure distribution that was lower at the bottom and higher at the middle nodes, accompanied by a higher flow velocity at the base. This confirmed the viability of CFD as a powerful tool for simulating complex transport mechanisms of assimilates in greenhouse tomatoes. Similarly, He et al. [49] developed a two-dimensional transient CFD model to optimize the ventilation structure of a removable rear-wall greenhouse. Their findings indicated that a 1.4 m ventilation opening significantly improved the thermal environment by enhancing airflow exchange, reducing the average indoor temperature by 1.7 °C, and achieving a maximum reduction of 5.8 °C under extreme conditions. Furthermore, Li et al. [112] proposed an interactive optimization framework in which the spatial influences of environmental parameters (air temperature, airflow, and $C{O}_2$ concentration) were explicitly considered through CFD simulations. By enabling data exchange between the optimization algorithm and CFD, this framework derived optimal environmental settings for greenhouse crop growth. CFD models excel in capturing spatial heterogeneity and flow dynamics, making them particularly suitable for design and structural optimization tasks where detailed spatial resolution is critical. these approaches are most appropriate when: (1) high spatial resolution is essential; (2) adequate computing resources are available; (3) the application involves design or structural optimization rather than real-time control; and (4) detailed physical understanding of flow phenomena is required.

3.1.3. Data Driven Modeling

With the rapid development of intelligent techniques, data-driven models have gained increasing attention due to their independence from prior mechanistic knowledge. The ARX model [113], characterized by its simple linear structure and convenient online identification, is suitable for short-term greenhouse temperature prediction. However, it shows limited capacity to capture nonlinear processes such as $C{O}_2$–transpiration coupling. In contrast, fuzzy models [114] and neural networks [115,116] overcome these limitations through nonlinear mappings, thereby offering greater flexibility in modeling complex relationships.

Recently, Jung et al. [117] used LSTM to predict tomato transpiration and humidity, achieving a 12% reduction in standard error compared with the Stanghellini model. They further developed a CNN-LSTM hybrid, where CNN extracted spatial correlations among variables and LSTM captured temporal dynamics, improving humidity prediction by 18% over standalone RNN-LSTM [80]. Similarly, Liu et al. [5] synthesized 61 studies, confirming that LSTM reduced prediction errors by 52.1% and 23.6% compared with RNN and GRU, respectively, while attention-enhanced LSTMs mitigated long-horizon accuracy decay. Lin et al. [118] used an LSTM-Sigmoid hybrid to predict 30-min internal temperatures solely from external climate inputs (R² = 0.962), then coupled the outputs with Logistic and Gompertz models for crop growth simulation, demonstrating the potential of lightweight greenhouse modeling with reduced sensor dependence.

A systematic study by El Aloui et al. [51] showed that data-driven models, particularly ANN and SVM, significantly outperform CFD methods in greenhouse microclimate prediction. For both temperature and humidity, these models achieved correlation coefficients R² exceeding 0.98, while reducing computation time from the order of 48 h required by CFD models to mere minutes.

To reconcile the accuracy-computational cost trade-off, model reduction techniques have emerged as an innovative pathway [25,119]. Our research group applied proper orthogonal decomposition (POD) to extract dominant features of multiple climate parameters from offline CFD simulations. Through multidimensional interpolation, the reconstructed low-order model provided environmental responses with high spatial resolution comparable to CFD while achieving response times on the order of seconds, boosting NSGA-II optimization efficiency by 88.09% [120]. Further, a rolling horizon control framework was designed for crop growth regulation, combining the POD-based climate model with Henten’s crop growth model [102]. By dynamically optimizing shading ratios and fan speed sequences, the method achieved a 43% increase in crop yield and a 13.8% reduction in energy consumption compared with traditional on/off strategy [121]. The schematic diagram of this approach is shown in Figure 12. Data-driven models are most effective when: (1) large, high-quality datasets are available; (2) medium to high computational resources can be allocated; (3) real-time or near-real-time predictions are required; and (4) the application involves complex, nonlinear relationships that are difficult to model mechanistically. However, they require substantial training data and computational resources, and their black-box nature limits interpretability.

We categorize greenhouse environment modeling methods over the past two decades in

Table 6. Summary of greenhouse environmental modeling approaches.

Type	References	Key Features	Limitations
Mechanistic models	[103,104,105,106,107]	Thermodynamic and physiological basis; dynamic equations; calibration by evolutionary algorithms.	Require many parameters; difficult calibration; heavy computation.
CFD simulation	[49,108,109,110,111,112]	Spatial distribution of airflow, temperature, $C{O}_2$; ventilation and transport analysis.	Very high computation; not suitable for real-time.
Data-driven models	[5,51,80,114,116,117,118]	No prior knowledge; ANN, fuzzy logic, LSTM, CNN-LSTM; accurate prediction.	Depend on data; weak generalization; ignore spatial effects.
Model reduction	[119,120,121]	POD from CFD; fast response; combined with crop models.	Approximation errors; need prior CFD data.

, providing a comprehensive comparison of their respective domains of applicability, computational requirements, and validation contexts.

3.2. Crop Growth Model

Crop growth modeling has undergone several decades of development and has evolved into a relatively comprehensive system, playing a crucial role in agricultural greenhouse production and resource management.

The mechanism model of crop growth is based on the laws of energy/mass conservation, and quantitatively describes key physiological and ecological processes such as photosynthesis, respiration, and phenological development through mathematical formulas. Early studies were mainly concentrated on canopy photosynthesis [122], and several models such as ELCROS, BACROS, and SUCROS were developed since 1970s [123,124,125,126,127]. In 1990s, the DSSAT system integrated the CERES and CROPGRO series of models [128,129,130], enabling simulation of the complete crop development cycle together with soil water and nitrogen dynamics. Meanwhile, APSIM in emphasized dynamic soil characterization and incorporated multiple crop and soil modules [131]. These models, strongly mechanistic in nature, established the theoretical foundation for simulating the interactions between crop growth and the environment.

Recent efforts highlight the transition from empirical formulations toward mechanism–data hybrid models, in which crop growth laws serve as structural constraints while intelligent methods capture nonlinearities and hidden processes [132]. For instance, Ali et al. [133] developed a CFD-based crop sub-model that coupled a porous medium framework with the Penman–Monteith transpiration equation, enabling dynamic simulation of microclimate–plant interactions and providing a digital twin platform for embedding machine learning algorithms. Similarly, Nomura et al. [134] established a coupled framework for light distribution, photosynthesis, and transpiration, integrating Monte Carlo ray tracing with leaf energy balance models and time-series canopy image analysis to non-destructively monitor leaf area index (LAI). In the applications of remote sensing data, Huang et al. [135] combined ensemble Kalman filtering (EnKF) and four-dimensional variational (4DVar) methods with Sentinel-2 satellite data within the WOFOST model, and further integrated LSTM networks for temporal simulation, resulting in improved modeling performance under stress conditions. Zare et al. [136] employed a particle filtering technique to assimilate remote sensing data into the PILOTE crop model, achieving real-time forecasts of winter wheat yields under uncertain weather and management conditions. More recently, Démoulin et al. [137] demonstrated the integration of a dynamic maize growth model with a 3D radiative transfer model (DART), which significantly enhanced the retrieval of canopy and leaf traits across growth stages using hyperspectral data.

Localized model calibration and parameter estimation are critical for crop growth modeling in controlled environments [138,139,140,141]. Acquah et al. [8] systematically validated the Stanghellini evapotranspiration (ET) model in naturally ventilated Venlo-type greenhouses in northeast China, confirming its applicability for tomato crops and identifying solar radiation and air temperature as key predictors. Building on this, Li et al. [84] refined the FAO-56 model by introducing a canopy height–wind speed correction formula and applied Bayesian optimization for adaptive parameter updating, thereby improving ET estimation accuracy. Deep reinforcement learning is beginning to be tested for the joint optimization of crop growth and greenhouse environmental control [142], with initial results showing improved energy–photosynthesis trade-offs. In addition, Graefe et al. [143] introduced Gaussian process regression (GPR) to build a three-dimensional light-response model incorporating leaf age effects. Applied to vertical tomato cultivation, this reduced V_cmax and J_max prediction errors by 13–17%, successfully quantifying the coupling between leaf senescence and light adaptation history. Chen et al. [144] employed an improved ResNet-50 convolutional neural network to extract the biodegradable mulch disintegration area index (DAI) in real time, and coupled it with soil temperature and moisture data to model dynamic interactions. This multimodal approach reduced yield prediction errors by 28.6%. Agricultural robotic platforms such as GRoW [145], equipped with LiDAR and multispectral sensors, provide high-throughput phenotyping data to enrich state diagnosis and alleviate the limitations of manual observations. Qin et al. [146] demonstrated that UAV-based pesticide spraying alters canopy microenvironments through droplet size distribution, thereby affecting pest habitats. This finding underscores the potential for integrating agronomic interventions into crop growth models.

4. Discussion

This paper reviews the intelligent control and related AI technologies applied in greenhouse systems in recent years. A brief discussion is provided in four aspects as follows.

(1) Although hierarchical greenhouse control framework is widely recognized as a reasonable solution for crop growth control, traditional control methods are still widely used and have reliable performance, especially in feed back control of single environmental variable. In terms of controlled variables, temperature is the most direct variable among environmental parameters. Deviation based PID or other traditional methods (such as fuzzy control) can be used to regulate temperature, provided that the ideal value for each cultivation stage of crop growth is known or can be provided through expert experience or crop growth models. Greenhouse environmental control has evolved from conventional strategies toward increasingly intelligent and adaptive methods. However, their inherent limitations in handling nonlinearities, multivariable couplings, and time delays make them insufficient for complex greenhouse dynamics. Recent enhancements such as self-tuning, fuzzy adaptation, and hybrid PID-fuzzy approaches have partly mitigated these shortcomings, yet their performance is often restricted by rule-base complexity or limited generalization capacity. Through embedded dynamic model, MPC can handle time delays and nonlinear characteristics in greenhouse environment, optimize multiple control variables (such as temperature, humidity, light intensity, etc.) simultaneously, and coordinate the relationships between variables through global optimization algorithms. For complex greenhouse systems, establishing such a dynamic model is the key issue to the successful application of MPC [52,53].

(2) Intelligent methods could significantly advanced the adaptability when faced with more complex greenhouse system regulation tasks. Artificial neural networks, including BP, Elman, and advanced recurrent structures, have emerged as powerful predictive models capable of capturing nonlinearities and temporal dependencies. Their use as state estimators or controller surrogates has improved prediction accuracy, shortened regulation times, and enhanced resilience against disturbances. In greenhouse system applications, ANN can serve as a predictor or state estimator for greenhouse environmental variables, as well as be used to directly design environmental controllers. When designing a neural network controller, it is possible to directly obtain an inverse model controller by training ANN to simulate the inverse dynamics of controlled object. But this approach heavily relies on the accuracy of the inverse model and has poor robustness. Usually, two neural networks need to be designed: the identification network and the controller network. The former approximates the forward dynamics model of the controlled object for predicting system behavior. The latter adjusts its own weights based on the system model information provided by the former model, calculates the controlled variable u so that the outputs tracking the ideal signal of the system, such as [72,74]. In terms of neural networks selection, BP network has a simple structure and is easy to implement; RNN and LSTM are more suitable for predicting and controlling time series data. Among them, LSTM is a special and more complex variant of RNN. Through a carefully designed “gating mechanism”, LSTM can effectively learn and remember long-term dependencies. In time series tasks involving long sequences, the performance of LSTM far exceeds that of traditional RNN. It can be foreseen that LSTM has more advantages in environmental regulation of the entire growth cycle of greenhouse crops. Nonetheless, data requirements, interpretability issues, and the risk of overfitting remain notable challenges for all kinds of data-driven methods.

(3) If intelligent control technologies such as neural networks are supplements and improvements to existing hierarchical greenhouse environment control frameworks, then RL provides a completely new control manner. Greenhouse environmental control is a complex decision-making problem, with the goal of regulating various actuators (such as heaters, ventilation windows, sunshades, $C{O}_2$ injectors, etc.) to create the optimal environment for crop growth (temperature, humidity, lighting, $C{O}_2$ concentration) while minimizing energy and water resource costs. This is precisely the field where RL excels [147]. RL’s capacity for end-to-end optimization allows simultaneous coordination of multiple variables and long-horizon decision-making, directly linking environmental regulation to crop yield and energy consumption. When combined with robust optimization or model-based simulators, RL exhibits strong potential in uncertain or extreme operating conditions. The mainstream solution is “Train in simulation and then deploy in reality”. Currently, multiple articles have introduced RL for optimal control of multiple variables in greenhouses [95,96,97,98]. These studies demonstrate that RL-based control, especially when integrated with deep learning techniques, provides a highly promising solution to the multi-variable and multi-objective coordinated optimization problem of greenhouse systems, and drives greenhouse control toward fully autonomous intelligent decision-making. Despite these advances, practical deployment of RL faces challenges such as computational demand, sample efficiency, and ensuring stability guarantees.

(4) In terms of greenhouse systems’ modeling, intelligent technologies also show great performances and application potential. For parameter identification and tuning of mechanism models, evolutionary algorithms utilize their stochastic and global optimization properties to find the optimal parameters and improve model accuracy effectively [106,107]. Although data-driven modeling is a “black box” approach lacking interpretability and heavily relying on on-site data, it can achieve more accurate approximation performance than mechanistic models in specific application scenarios if field data set is sufficient and high quality. Various ANNs and their hybrid forms have been successfully applied. Among them, due to its ability of remembering long-term dependencies, LSTM has been widely used for modeling and predicting greenhouse environments and crop growth, such as [5,80,117,118]. The summary of data-driven models for greenhouse environmental modeling is presented in

Table 7. Summary of data-driven models for greenhouse environmental modeling.

Model Type	References	Key Features	Limitations
Fuzzy models	[114]	Nonlinear mapping; better than ARX.	Rule design subjective.
ANN	[51,115,116]	Accurate climate prediction; faster than CFD.	Data-dependent; weak generalization.
LSTM/RNN	[5,80,117]	Capture dynamics; LSTM outperforms RNN/GRU; CNN-LSTM improves humidity prediction.	Long-horizon decay; data hungry.
Hybrid	[118]	LSTM-Sigmoid + growth models; accurate with fewer sensors.	Higher complexity; low interpretability.
SVM	[51]	High accuracy; efficient vs CFD.	Sensitive to kernel; no spatial info.

.

In terms of reducing modeling complexity and computational cost, the model reduction method is an innovative way. Through offline-online strategy, dominant feature vectors of a dynamic model could be extracted by feature decomposition, enabling low dimensional reconstruction with minimal energy loss. Based on the reduced surrogate model, optimal control framework could be performed more efficiently [120,121].

5. Future Challenges and Development Trends

(1) Robustness and generalization Most current machine learning models require high-quality datasets to drive them. Different greenhouse field data can lead to deterioration in accuracy or even model failure. Future research should aim to reduce the dependence of models on large volumes of labeled data while ensuring stable performance across diverse conditions. The fusion of meta learning or transfer learning is a possible way to reduce on-site data dependence, and improve the robustness and generalization performance of data-driven models.

(2) Sustainability and energy management Although current greenhouse management systems have identified energy conservation as an optimization goal, with the continuous development of microgrid technology, the regulation of environmental variables such as sunlight and temperature should be coordinated with renewable energy systems more tightly. For such cross-system optimization, reinforcement learning offers natural advantages, but existing studies remain preliminary.

(3) From “standalone intelligence” to “cluster intelligence” Future research will focus more on multi-greenhouse collaborative control to improve overall production efficiency. Through the Internet of Things and cloud computing, centralized monitoring and intelligent decision-making of multiple greenhouses can be achieved in a more efficient way.

6. Conclusions

The complexity and variability of agricultural greenhouse systems pose significant challenges to achieving efficient and optimal crop production. Therefore, intelligent control strategies with high accuracy and adaptability are of great importance. This paper reviewed relevant studies in the past fifteen years on greenhouse system control technologies. The review highlighted the following aspects: (1) Conventional control approaches such as PID, fuzzy, and model predictive control were summarized, and their limitations in handling nonlinear and strongly coupled variables were analyzed. PID-based methods still account for a considerable share of studies, especially in temperature and irrigation regulation. In general, their stabilization times were reduced by more than 70% in temperature control, and irrigation errors were kept within ±2–5% in field trials. However, their performance degrades in multivariable or highly nonlinear scenarios, where advanced methods become necessary. (2) A comprehensive survey of intelligent control methods was presented, with particular focus on neural network-based models including BP networks, RNN, and LSTM, as well as adaptive control, feedback linearization, event-based control, and RL-based strategies. Their respective advantages and application scenarios were compared. Across the surveyed studies (Table 4), NN-based approaches typically reduced prediction errors by 20–50% compared with conventional baselines, while RL-based strategies achieved more holistic improvements, including up to 50–60% energy savings and measurable yield increases in simulation and pilot-scale trials. These quantitative improvements highlight the potential of intelligent methods to significantly enhance greenhouse productivity and resource efficiency. (3) Advanced modeling techniques for greenhouse environments were also discussed, including mechanistic models, CFD simulations, and data-driven modeling approaches, with emphasis on the trade-off between accuracy, interpretability, and computational cost.

Finally, possible research directions were proposed, emphasizing the need for real-time adaptability, sustainability, and multi-greenhouse collaborative intelligence to meet future food production challenges.