Modelos heterogêneos para a previsão de safras e qualidades de cultivo na indústria sucroenergética

Abstract

Agriculture is an important sector for the Brazilian and world economy and society. The agroindustries together with small farmers are responsible for placing the country among the largest producers of this sector in the world. An important representative of the Brazilian agro-industry are the sugar-energy plants, responsible for the planting, cultivation and production of sugarcane by-products. Sugarcane accounts for 80 % of all sugar produced, Brazil being the largest producer of sugarcane in the world. Agriculture has been gaining a lot of strength, productivity and innovation, especially when aligned with recent technology concepts. Precision farming makes use of innovative software and hardware resources to evaluate and monitor crop conditions. With this, it generates a large set of data prepared for analysis and later serve for decision making. However, agroindustries usually use large areas for the planting of different types of monocultures. The size of these areas usually poses major problems for monitoring, pest control, the environment and planting. In order to maximize production while seeking to minimize environmental impacts, agribusiness increasingly seeks to invest in technologies and research that assist in the analysis of various data collected throughout the harvests. These data need to be processed to extract relevant information. Given this need, this work uses CRISP-DM as a process for data mining, which is a very useful tool as part of the definition of the project methodology. With the realized surveys and literature review, it is realized that it is necessary to make predictions and analyzes of future behaviors from modeled objects. However, the quality of the answers of any proposed model depends on the precision of the computational structure and the data that feed the model. The problematic raised is a complex system, that is, a system that is composed of numerous elements that interact, so that the aggregate behavior can not be inferred from the behavior of the constituent units alone, thus falling within the dynamics of cultivating a covering different variables of interest, such as temperature, luminosity, soil, among others. In order to understand the problematic and development of the solution proposals, the methodology of Tookit HCD was used, a process with principles focused on immersion to understand the problems and development of solutions in line with the needs of the end users. This dissertation discusses and suggests a pilot model for implementing an infrastructure for collection, storage, processing and visualization of plantation data through the use of the internet of things and mobile devices. The contribution is based on the presentation of two stochastic models for the community, focused on the prediction of harvests, quality indexes and cultivation scenarios of the different stages of sugarcane growth. The models obtained positive and promising results in the simulations. Data from the São José Agroindustrial Plant was used in a study area of 15 hectares. The first model presented, based on the use of the Monte Carlo method in Markov chains, obtained good results in the prediction of crops and quality indices of sugar cane. The model obtained in experiments hypothesis tests with p-values of 0.8754 and kappa coefficients of 0.68. The other model is based on stochastic cellular automata, which aims to simulate georeferenced scenarios of the plantation, also classifying regions as good, bad or medium. The model obtained in experiments a p-value of 0.8635 in the hypothesis test and a kappa coefficient of 0.71. In both models the p-values and kappas indicate a positive relationship between the model results and the base data. Investment in agricultural technologies and innovations is essential to optimize production in this sector, as well as reduce costs while preserving the environment.