Analysis of approaches to creating a «Smart Greenhouse» system based on a neural network

Cherepenin Valentin Anatolyevich ORCID: 0000-0002-6310-1939 Postgraduate student, Department of Information and Measurement Systems and Technologies, South Russian State Polytechnic University (NPI) named after M.I. Platova 132 Prosveshcheniya str., Novocherkassk, Rostov region, 346428, Russia cherept2@gmail.com Other publications by this author

Katsupeev Andrei Aleksandrovich PhD in Technical Science Associate Professor, Department of Information and Measurement Systems and Technologies, South Russian State Polytechnic University (NPI) named after M.I. Platova 132 Prosveshcheniya str., Novocherkassk, Rostov region, 346428, Russia andreykatsupeev@gmail.com

Introduction

With the development of Internet of Things (IoT) and artificial intelligence (AI) technologies, Smart Greenhouse systems are becoming an important direction in modern agriculture. Efficient use of resources, optimal climate control and automation of plant growing processes provide agricultural enterprises with new opportunities to increase yields and reduce costs. This article provides an overview of existing approaches to creating a Smart Greenhouse system, with an emphasis on the use of neural networks. The research aims to identify the key aspects and benefits that neural networks provide in the context of greenhouse climate management. The methods of data collection and analysis, the introduction of sensor technologies, as well as information processing algorithms using neural networks to optimize plant growth conditions are considered. This review provides a comprehensive understanding of the current state of Smart Greenhouse technologies and highlights the prospects for further research in this area.

Research and comparison of climate control methods in greenhouses

The climate in the greenhouse is based on the environmental conditions necessary for plants to live. The greenhouse microclimate is very complex, nonlinear, multiparametric, and depends on a set of internal and external factors, including meteorological factors such as humidity and ambient temperature, solar radiation intensity, wind speed and direction.

The components of a greenhouse are internal factors, various crops, greenhouse components and elements such as fogging, heating, ventilation systems, soil types, etc. The different description of the greenhouse climate is explained in two approaches: one is based on mass energy consumption equations that describe the process. The other uses a system identification approach consisting of examining both the input and output data of the process ^[1]. Thus, these approaches are designed to solve all microclimate problems based on modern methods of adaptive and nonlinear systems. In order to achieve predetermined and optimal results, climate control in greenhouses creates a favorable growing environment. Several management strategies and methods have been proposed to manage the greenhouse environment, such as adaptive management, predictive management, fuzzy control logic, reliable management, nonlinear feedback control, and optimal management. Many studies have been conducted in which energy savings are being sought, where energy collection optimization methods, physical models and computational fluid dynamics methods are used to predict the microclimate of greenhouses ^{[2, 3]}.

These forecasting methods are divided into two groups:

            1. The physical method.

            2. The black box method.

The first method is based on mathematical theory, which is necessary to regulate a huge number of parameters, and the calculation of these parameters is difficult. The black box method is based on modern computing technologies, which do not always ensure convergence to an optimal solution and even easily undergo partial optimization.

On the other hand, neural networks, unlike programmed ones, learn to implement all patterns. These greenhouse systems are very appropriate for reflecting knowledge that cannot be programmed to represent nonlinear phenomena ^[4].

Training and application of neural networks in agriculture

Since greenhouses are unchangeable over time, nonlinear, and strongly interconnected, a lot of research is underway to select artificial neural networks (hereinafter ANN) for modeling, optimizing, predicting, and controlling all these processes. This section covers all of ANN's various applications and research in greenhouse technology. The development of this type of model will expand the possibilities of its application in the future and combine with technologies that are in demand in intelligent agriculture. Figure 1 shows the classification of models of various greenhouses.

Fig. 1. – Classification of greenhouse models

ANN is a machine learning algorithm based on the analysis and understanding of a human neuron. This is a computational model with biological influence, consisting of processing elements (neurons) and the relationships between them with coefficients, usually called weights ^[5].

Neurons receive input data in the form of a pulse. Some measures used to describe the activity of neurons are the peak rate generated over time and the average peak generation in several runs. A neuron is recognized by the rate at which it produces peaks. With the help of adaptive synaptic weights, neurons are connected to other neurons of the previous layer. Knowledge is usually stored in a set of connection weights. The learning process identifies a suitable learning method when these compound weights change accordingly. These training methods contain input data for the network and generate the desired result by changing the weights to get the ideal result. As the training progresses, the weight methods will have relevant information compared to the previous training, since it contains all the redundant and meaningless information ^[6].

In neural networks, learning is an important part. Training a neural network model is the process of adjusting the weights and offsets of the network so that it can accurately predict the output for a given input. This is done by repeatedly presenting the network with examples of input and output data, and then adjusting the weights and offsets so that the network forecasts become more accurate. The learning process is repeated many times until the model converges, which means that the error is no longer significantly reduced. The number of repetitions of the learning process is called the number of epochs. This process will determine the connection between input and output, as it requires the most accurate predictive calculation. This learning process is divided into two categories:

1. Under supervision;

2. Unattended.

Supervised learning perceives the expected results and labels the data. In unsupervised learning, it is not necessary to perceive data, since learning is carried out by determining the location of the data representation and internal structures.

The network structure depends on the type of task being described, the complexity of the system, and the learning process. Table 1 shows the various types of neural models, describes in detail the input and output variables that were used, the network architecture, activation functions that were used by the network or each network layer, and the algorithm used in the training process ^[7].

                                                         Table No. 1. Types of neural network models for forecasting

Neural network training is reduced to minimizing the error function by adjusting the weights of synoptic connections between neurons. The error function refers to the difference between the received response and the desired one. For neurons of the output layer, their actual and desired output values are known. Therefore, setting the connection weights for such neurons is relatively simple ^[8]. However, for the neurons of the previous layers, the setting is not so obvious. For a long time, there was no known algorithm for spreading the error across hidden layers. The error value is determined by the following formula:

,                                                                        

where Ep is the magnitude of the error function for image p; tpj is the desired output of neuron j for image p; ypj is the activated output of neuron j for image p.

The teacher learning method involves reducing the cost function, which collects errors between the desired output and the actual output for the given input data. To reduce this cost function, several methods are used.  The BP algorithm is one of the most commonly used to obtain results in single-layer and multi-layer networks. The algorithm is designed for ANN self-organization and is derived according to Hebb's law. The revealed learning methods allowed the development of advanced algorithms such as self-organizing maps (hereinafter S.O.M.) and self-organizing tree algorithm (hereinafter SOTA). Based on unsupervised data, these are time series clustering algorithms. The S.O.M. tool is used for data visualization and clustering. The main disadvantage of this tool is that the user needs to choose the size of the map, which leads to numerous experiments with maps of different sizes in an attempt to get the optimal result. This process can be quite slow. A more accurate classification of patterns in the last SOTA layer will allow classifying a group of patterns separated from each other at the initial stages. Using SOTA methods, neural networks can be built automatically from the available training data.

According to these methods of application and processing of tools, climate-related variables for the greenhouse are crucial. Some parameters, such as calculating speed, predicting accuracy behavior, and managing variables of various elements, remain a huge challenge. Using some nonlinear methods, artificial neural networks largely solve these problems.

Optimization of the microclimate in greenhouses using neural networks

To achieve controlled agricultural production, greenhouses are nonlinear and complex systems. These systems have internal factors, a composite dynamic impulse of external factors, and control mechanisms. Based on the microclimate and management of greenhouse crops, it acts as a central element and an essential part of the system. This system usually takes into account some general issues that are more important reactions of the microclimate, since it represents the complex and diversity of crops in the greenhouse. The greenhouse system mainly focuses on the set of the environment that affects the development and growth of crops. The microclimate system in greenhouses has attracted close attention in recent years due to its significant contribution to improving yields. Statistics, engineering and artificial intelligence have been introduced into the system to predict various elements, such as the amount of CO2, temperature and relative humidity.

The greenhouse system uses conventional control to ensure the desired performance, but this control may be unsatisfactory. Recently, the topic of using neural networks to control the microclimate has been of interest. To reflect the nonlinear characteristics, the neural network provides all reliable models for the greenhouse, since it is difficult to solve using traditional methods, since reliable models do not require any prior knowledge of the system, the real-time model works quite satisfactorily for dynamic systems ^[9].

The most important attributes in the greenhouse microclimate system are humidity and temperature. Complex gas exchange can occur in the environment, as well as the interaction between mass, heat and internal air, as well as some elements of the greenhouse from the outside. Building a reliable model is a very difficult task, since it must fulfill all parameters, mathematical functions and transformation functions ^[10]. When we build a model using ANN, we see great opportunities to display all non-linear functions to create numerous production process systems. To build a model, you need to design a network of the system, such as humidity and air temperature in greenhouses, which are considered as output data; this is caused by some facts. For the model, the most difficult thing is to set the input data in order to better understand the system. To consider the input variable in the system, the following considerations must be followed:

        1. The nature of the physical dependence of the input and output.

        2. Correlation between inputs and outputs.

        3. The range of input variables.

The input variables of the system are internal relative humidity, internal temperature, internal atmospheric pressure, external temperature, external atmospheric pressure, external temperature, wind speed and direction. The output variable of the system is CO 2. The transfer function is used in all layers. 0.97 coefficient was obtained using the forecast of the concentration of CO 2. An artificial neural network must be trained in accordance with all possible situations and measurable objects in order to obtain a unique result with limited data ^[11].

A database is required for the proper functioning of a neural network. This section describes many models in which these artificial neural networks are used to predict the microclimate of a greenhouse system, and network optimization is related to its structure and type. Since the model requires training data, data collection is one of the key points that is equally important to the system for efficient output. There will be three general steps in each database, such as testing, training, and validation. These phases must be taken into account in each neural network. The Backbox model requires a huge amount of data in ANN ^{[12, 13]}. The optional sum in the black box model is a trial and error method that selects the hidden layer and neurons. Since we consider a huge amount of data to obtain an accurate conclusion, this leads to a decrease in the efficiency of the process. Some methods are applied to a model with huge data that can simplify and optimize the N.N. learning process. A 12-month database is ideal for ANN to predict a greenhouse neural network that has a huge amount of data. The model was selected for 29 days: 22 days of the training phase and 7 days of the testing phase. The model was considered phases divided into 15% for validation, 15% for testing, and 70% for network training. Based on the model and requirements, it is assumed that the period gives an accurate result. Fig. 2 shows the forecast of the greenhouse microclimate in an artificial neural network.

Fig. 2. - Forecast of greenhouse microclimate in an artificial neural network

Conclusion

Unlike physical models, ANN takes only a few minutes to complete the indoor climate forecast, given that there are many unknown factors involved that cannot be studied using physical models. The combination of ANN and physical models would make it possible to better predict the microclimate, however, the design of this hybrid network has been little studied and, therefore, this network should be studied. In other applications of neural networks in greenhouses, the assessment and prediction of epidemics and diseases of crops can be carried out using technologies such as image analysis combined with deep learning models such as CNN. Similarly, microclimate prediction may be feasible using these methods, since these methods can process a larger amount of information compared to traditional ANN models. The role of ANN is to develop predictive models that take advantage of the information generated and manage it.