Greenhouse environmental monitoring and control system based on improved fuzzy PID and neural network algorithms

3.1 Design of a GEMC system based on improved fuzzy PID controllers

The fuzzy PID controller is a controller that combines PID control and fuzzy logic control [18]. PID control is a common control strategy that adjusts the proportional, integral, and differential parameters of the system to control its output [19]. However, PID controllers may have poor control performance for nonlinear systems or systems with constantly changing parameters. Fuzzy logic control is a control strategy based on fuzzy set theory, which processes complex, uncertain, or nonlinear problems using human language or logic by fuzzifying input variables. The fuzzy PID controller combines the advantages of both and dynamically adjusts the parameters of the PID controller through fuzzy logic, enabling the controller to better cope with nonlinear systems or systems with parameter changes [20]. The fuzzy PID algorithm has advantages such as strong adaptability, self-learning ability, robustness, and the ability to optimize control decisions and reduce system oscillations. At the same time, it also has limitations such as complex design, high computational cost, difficulty in adjusting parameters, and over-fitting. To address these limitations, research is being conducted using neural network algorithms for parameter tuning. The neural network algorithm is an algorithm that simulates the working mode of human brain neurons, which can recognize and predict complex patterns by learning a large amount of data. In the GEMC system, neural networks can be used to predict environmental conditions or act as controllers to adjust the environment to achieve optimal conditions. The mathematical expression for the proportional controller, i.e., the P controller, in the fuzzy PID controller is shown in equation (1).

In equation (1), U P represents the P controller, k P represents the proportional gain, e ( t ) represents the error of the system, and t represents the time. The expression for the proportional integral controller, i.e., the PI controller, is given as equation (2).

In equation (2), U PI represents the PI controller, T I represents the time of integral control in the system, and k I represents the integral gain. The PID controller is given in equation (3).

In equation (3), U PID represents the PID controller, T D represents the time of differential control in the system, and k D represents the differential gain. The principle of the fuzzy PID controller is presented in Figure 1.

Figure 1

Principle of the fuzzy PID controller. Source: This figure has been created by author.

In Figure 1, the operations of the fuzzy PID controller mainly include fuzzification, fuzzy inference, and deblurring [21]. This process requires interpretation and decision-making of input values. However, in GhE, the measurement accuracy of sensors and the interference of environmental factors may affect the accuracy and stability of temperature and humidity control. Traditional fuzzy PID controllers may not be able to meet the precise control requirements of these environments. Therefore, this study improves the traditional fuzzy PID controller by applying the least squares support vector machine (LS-SVM) reverse identification models and proposes an improved fuzzy PID controller. The improved controller has an inverse composite control effect to enhance the effectiveness of GhE control. The structure of the improved fuzzy PID controller is displayed in Figure 2.

Figure 2

Improving the structure of the fuzzy PID controller. Source: This figure has been created by author.

O 1 and O 2 in Figure 2 represent the output temperature and humidity, respectively. I 1 and I 2 correspond to the inputs of the temperature and humidity fuzzy controller for the greenhouse, respectively. This improved fuzzy PID controller mainly consists of a PID controller, a fuzzy controller, and a scaling factor determiner [22]. This structure utilizes a variable universe fuzzy controller to adjust the temperature regulation u 1 of the heating water pipe and the speed regulation u 2 of the fan, in order to achieve temperature and humidity regulation. Then, the system will output the actual adjustment of water pipe temperature U 1 and the adjustment of fan speed U 2 . These actual output data will then be fed into a fuzzy controller for GhE regulation. The fuzzification calculation formula of the fuzzy controller is presented as equation (4).

In equation (4), W represents the fuzzification result of environmental parameters such as temperature and humidity, L represents the value of the domain, and w L and w H represent the maximum and minimum values of environmental parameters during detection, respectively. The expression of the scaling factor is given in equation (5).

In equation (5), δ represents the scaling factor, λ represents the adjustment amount of the expansion and contraction amplitude, K represents the scaling factor, and m represents the measurement of deviation. The GEMC design based on an improved fuzzy PID controller is presented in Figure 3.

Figure 3

Design of the GEMC system based on an improved fuzzy PID controller. Source: This figure has been created by author.

In Figure 3, the design of GEMC based on an improved fuzzy PID controller mainly consists of the solar power supply module, communication module, data acquisition module, and terminal node module. The solar power supply module converts solar energy into electrical energy through solar panels and is usually used in conjunction with batteries to ensure that the system can still operate in the absence of sunlight. This module can provide clean and renewable power for all components, reduce system operating costs, and increase its adaptability to areas far from the power grid. The communication module enables the system to remotely transmit data and receive control commands, thereby enabling remote monitoring and management and reducing the need for manual intervention. The data acquisition module is composed of sensors used to measure environmental parameters such as temperature, humidity, light intensity, and CO₂ concentration. The data collected by this module can be used for monitoring GhEs. The terminal node module is composed of actuators, and the execution controller can adjust the environmental conditions inside the greenhouse based on the commands received from the central processor. In this configuration, the environmental data collected by the sensors are sent to the central processor for processing, and then through the coordinator played by the micro-controller the data flowed to the communication module to perform communication tasks. All of these modules are powered by solar energy, reflecting the concept of environmental protection and sustainable development.

3.2 Design of a GhE prediction model for the PID control algorithm based on a BPNN

GhE is a complex coupled system composed of multiple factors such as temperature, humidity, and light [23]. Among them, temperature and humidity are key factors that are influenced by a combination of other environmental factors. Therefore, to optimize control, it is necessary to decouple temperature and humidity to ensure that each output is only affected by one input. BPNNs have strong self-learning and adaptive abilities, making them suitable for nonlinear problems in multi-variable coupled systems. LS-SVM can approximate the inverse model of the original system and concatenate the inverse model before the original system to form a pseudolinear composite system. Therefore, this study applies a BPNN and LS-SVM for reverse identification of GhE. Furthermore, this study applies LS-SVM to the BPNN to improve the identification accuracy and efficiency. Figure 4 shows the BPNN process.

Figure 4

BPNN process. Source: This figure has been created by author.

In the BPNN process shown in Figure 4, environmental information is first collected and classified for training. The trained environmental factors are normalized for ease of processing and feature extraction. The network initiates the propagation of the input signal in a forward direction, calculates the output based on the input, and then performs backpropagation based on the output error. This process involves the adjustment of weights to reduce errors, thereby achieving self-learning and optimization of the network [24]. After completing an iteration, if the preset number of iterations or convergence conditions are met, the network will save the current weights and the model structure. Otherwise, the network will continue to iterate and optimize until the termination condition is met. In this way, the BPNN can optimize weight parameters, improve the accuracy of environmental factor identification and classification, and provide an accurate decision-making basis for GhE control. The expression for the output layer error is given as equation (6).

In equation (6), E represents the output error function, while d j and y j represent the actual output and expected output of neuron j , respectively. When the input quantity is a weight of 1, the output expression of the output layer is given as equation (7).

In equation (7), y i represents the actual output of neuron i and s j represents the input of neuron j in the output layer. The expression of the error signal at this time is given as equation (8).

In equation (8), δ j represents the error signal of neuron j . The expression for adjusting the weight parameters of the output layer is given as equation (9).

In equation (9), ω represents the weight, Δ ω represents the weight update, η is a preset constant, and ∂ represents the learning rate. To improve the convergence speed of the model, a momentum factor is introduced in the BPNN. The mathematical formula for the variation of momentum after the introduction of momentum factor is presented as equation (10).

In equation (10), Q ( t + 1 ) represents the amount of change at time t + 1 , β represents the momentum factor, and D t represents the negative gradient at time t . Figure 5 shows the PID control algorithm flow based on the BPNN.

Figure 5

PID control algorithm flow based on the BPNN. Source: This figure has been created by author.

As shown in Figure 5, the PID control algorithm based on the BPNN initializes the network structure and sets the output layer weights. Then, fuzzy inference is performed based on preset fuzzy rules by calculating errors and adaptively adjusting parameters and weights based on the errors. If the preset number of iterations has been completed, then the result is output and the algorithm ends. Otherwise, the algorithm returns to execute fuzzy reasoning, continuing to calculate and adjust until the desired number of iterations is reached. LS-SVM originates from traditional SVM and is a machine learning algorithm [25]. LS-SVM can obtain the optimal solution by minimizing the objective function [26]. In the reverse identification of GhE, this study uses LS-SVM to map input data to high-dimensional space to solve the problem of nonlinear datasets. The kernel function expression of LS-SVM is given in equation (11).

In equation (11), K represents the kernel function of LS-SVM, a and b represent input vectors and output vectors, respectively, and σ represents the kernel function parameter. The discriminant function expression constructed by LS-SVM is given as equation (12).

In equation (12), x represents the input sample and ϕ ( x ) represents the nonlinear mapping of sample x . The optimized minimum error expression is given as equation (13).

In equation (13), min S represents the minimum error, c represents the Lagrange multiplier, ξ represents the relaxation factor, and H represents the identity matrix. To convert nonlinear problems into a system of linear equations for solution, the transformed system of linear equations is given as equation (14).

In equation (14), Y represents the output sample matrix and R R T represents the square matrix of S × S . The expression of the LS-SVM regression function is given as equation (15).

In control systems, LS-SVM can be used to construct inverse models. Through the process of reverse identification, it is possible to gain a deeper understanding of the relationships between various environmental factors, thereby optimizing control strategies for GhEs and improving control effectiveness. In complex GhEs, there is a coupling relationship between various environmental factors. Reverse identification can help understand and reduce the impact of this coupling, so that each control variable is only affected by one input, thereby improving the efficiency and stability of control.

4.1 Experimental environment settings

In the experimental environment of GEMC, the main focus is on hardware configuration. The experiment used SHT75 sensors to monitor the temperature and humidity, as well as BH1750 light sensors to measure the light intensity. Simultaneously, a hot water pipeline system and a fan were selected to control the temperature and humidity separately. In terms of communication, Zigbee wireless technology was used as the communication device in the experiment to ensure the reliability and real-time performance of data transmission. In addition, the system also used solar power as the energy equipment for power supply. The overall architecture of the experimental environment consists of three phases. The initial phase of the process employs a front-end data acquisition module for the collection of greenhouse data. The second phase leverages an inverse model constructed by LS-SVM for the decoupling control of the GhE. The third phase utilizes Zigbee as a communication device to transmit the aforementioned data to the control execution module for operation and adjustment in accordance with the instructions of the control system. In the greenhouse control system, fans and heating pipes were used for room temperature control. The rated power selected for the fan was 100 W, and the rated speed was 1,420 rpm. As for the heating process, research is being conducted on using heating pipes to transport hot water to increase the indoor temperature. This study used MATLAB as the simulation software and selected Adam as the optimizer, with a set learning rate of 0.2 and a momentum factor of 0.02. The maximum number of iterations was set to 200. This study divided historical greenhouse data into training and testing sets in a 3:7 ratio. The experiment used temperature, humidity, and light sensors for data collection, and solar power supply was used to power the equipment. To achieve precise temperature control, heating water pipes and fans were used as control equipment for the heating and cooling system. Table 1 shows the specific hardware experimental environment.

4.2 Performance verification based on improved fuzzy PID and neural network algorithms

To evaluate the performance of the improved BPNN algorithm, the study compared its loss function values with traditional BPNN algorithms and GA, and the results are shown in Figure 6. Figure 6(a) shows that in the training set, the improved algorithm’s loss function value was reduced to 0.21, which was 40% lower than the traditional algorithm’s minimum loss function value of 0.35. Compared to the minimum loss function value of 0.53 in the GA, it was reduced by 60.37%. As shown in Figure 6(b), in the test set, the improved BPNN algorithm converged after 60 iterations, while the traditional algorithm and GA required nearly 110 and 80 iterations, respectively, to achieve convergence. In addition, the minimum loss function values of the improved algorithm, traditional algorithm, and GA were 0.32, 1.01, and 0.75, respectively. This meant that the improved algorithm reduced errors by 68.31 and 57.33%, respectively. In summary, the improved BPNN algorithm exhibited significantly improved performance. It not only showed fast convergence speed in the training and testing stages but also had significantly lower loss function values than traditional BPNN algorithms and GA, proving its significant optimization effect.

Figure 6

Comparison of loss function values for BPNN algorithms: (a) training set and (b) test set. Source: This figure has been created by author.

In Figure 7, to verify the performance of the improved fuzzy PID controller, this study compared and analyzed the simulation performance of the improved fuzzy PID controller and the preimproved controller in temperature and humidity control. In Figure 7(a), the solid line represents the expected temperature. The preimproved fuzzy PID controller only reached the stable expected temperature at around 1,600 s, while the improved fuzzy PID controller had already reached the stable expected temperature near 1,000 s, reducing the time by nearly 37.5%. In Figure 7(b), the time it took for the preimproved fuzzy PID controller to reach the expected humidity was about 1,700 s, while the improved controller had already reached the expected humidity at about 1,200 s, reducing the time by nearly 29.4%. Therefore, the improved fuzzy PID controller had a significant improvement in response time for temperature and humidity control compared to the previous controller.

Figure 7

Comparison of (a) temperature and (b) humidity simulation of the fuzzy PID controller before and after improvement. Source: This figure has been created by author.

To verify the performance of the improved fuzzy PID and BPNN algorithms, this study compared the improved GEMC with the unimproved system. The comparison of tracking performance of the control system before and after improvement is shown in Figure 8. The controller output of the unimproved system reached a convergence state at 58.7 s. The improved system only converged at 15.6 s, which shortened the response time by 73.42% compared to the unimproved system. The above data indicated that the improved control system had significantly better response efficiency than the unimproved system.

Figure 8

Comparison of the tracking performance of control systems before and after improvement. Source: This figure has been created by author.

To further validate the performance of improved fuzzy PID and BPNN, this study conducted simulation tests in the test set. During this process, the temperature of the heating water pipe and the wind speed of the fan were simulated. The identification results of the inverse system based on improved fuzzy PID and divine BPNN are shown in Figure 9. In Figure 9(a), the algorithm calculated a high degree of overlap between the temperature and the actual temperature. In the 40th set of data, the maximum error was observed, at which point the actual temperature was 58.04°C. The temperature calculated by the algorithm was 56.47°C, and the relative error value reached the maximum of 2.70%. As shown in Figure 9(b), there was also a significant overlap between the wind speed calculated by the algorithm and the actual wind speed, and the maximum error was observed in the 80th set of data. The actual wind speed at this time was 10.54 m/s, while the algorithm calculated wind speed was 9.03 m/s, and the relative error value at this time reached 14.32%. In summary, the improved fuzzy PID and BPNN had high identification ability in control systems.

Figure 9

Identification diagram of inverse systems – (a) heating water pipes and (b) wind speed for fans – based on improved fuzzy PID and neural network algorithms. Source: This figure has been created by author.

4.3 Verification of the actual effectiveness of GEMC

To verify the actual effectiveness of GEMC, this study compared the environmental predicted value (EPV) of the system with the actual measured value (AMV). Figure 10 shows a comparison between EPV and AMV. In Figure 10, the EPV and AMV of environmental factors such as temperature, humidity, and light almost completely overlapped, indicating that the model had extremely high predictive accuracy. As shown in Figure 10(a), the average EPV of temperature was 24.97°C, while the average AMV was 25.18°C. The EPV was slightly lower than the AMV, with a difference of only 0.83%, indicating that the temperature prediction was very close to the actual measurement results. As shown in Figure 10(b), the average EPV of relative humidity was 39.81%, while the average relative humidity of AMV was 42.03%. EPV decreased by 2.22% compared to AMV. Figure10(c) shows that the average intensity of EPV under illumination was 8331.1572 lux, compared to the actual measured average intensity of 8437.0157 lux. Compared to AMV, EPV decreased by 1.25%, indicating that the prediction of light intensity was also quite close to the actual measurement results. In summary, EPV was slightly lower than AMV overall, but the difference between the two was very small, indicating that the model had high prediction accuracy.

Figure 10

Comparison between EPVs and AMVs of (a) temperature, (b) humidity, and (c) illumination. Source: This figure has been created by author.

To further validate the effectiveness of GEMC, this study randomly selected five sets of environmental data for greenhouses and compared the environmental data monitored by the system with the measured data in detail. Table 2 presents a comparative analysis of environmental data. In Table 2, the error values of the system in temperature monitoring ranged from 0.1 to 1.3°C, with relative errors ranging from 0.35 to 5.04%. In terms of humidity monitoring, the error value ranged from −1.3 to 1.65%. As for lighting monitoring, the error value ranged from −102 to −8 lux, and the relative error was between −3.5 and −0.79%. Overall, this GEMC had a relatively high accuracy in measuring the temperature, humidity, and lighting, especially in terms of performance in measuring the temperature and humidity. However, further optimization may be needed in lighting measurement to reduce errors.

To verify the superiority of the GEMC system based on improved fuzzy PID and neural network algorithms, this control system was compared and analyzed with the most advanced greenhouse control methods. Advanced greenhouse control methods included heuristic algorithm-based greenhouse control systems, variable universe fuzzy control-based greenhouse control systems, and deep reinforcement learning-based greenhouse control systems [27,28]. The error comparison of environmental monitoring data from different greenhouse control systems is shown in Figure 11. As shown in Figure 11, from the perspective of temperature monitoring, the maximum relative error obtained by the control system studied in this article was 5.04%. The maximum relative errors of the greenhouse control system based on heuristic algorithms, variable universe fuzzy control, and deep reinforcement learning were 8.12, 10.01, and 9.54%, respectively. The research method reduced them by 3.08, 4.97, and 4.50%, respectively. From the perspective of humidity monitoring, the maximum relative error of the research method was −1.77%, which is still the smallest error compared to the maximum error values of 2.73, 4.42, and −2.60% of the other three methods. From the lighting data, the maximum error of the research method was −3.50%, which still had advantages compared to the maximum error values of the other three methods of −4.01, −3.49, and −4.40%. In summary, the GEMC system based on improved fuzzy PID and neural network algorithms studied had superior performance compared to other advanced greenhouse control methods.

Figure 11

Comparison of errors in environmental monitoring data of different greenhouse control systems: (a) temperature, (b) humidity, and (c) lighting data. Source: This figure has been created by author.