Inovação em completação de dados climáticos: métodos baseados em vizinhos, regressão linear e redes neurais de camadas densas

Abstract

This research aimed to develop and evaluate data completion methods for climatic databases. Following a literature review in the Web of Science database using the keywords "missing data," "data climate," "climate," and "imputation," a trend was identified in the use of machine learning (ML) methods for achieving more accurate results. Further research was conducted to survey scientific articles that use ML to understand the methods and define the research choices. The research started with an API that collects real-time data from INMET and already provides users with choices of stations and data periods to generate line graphs and box plots of climatic variables. This API was initiated in a previous study and finalized at the beginning of the current research to incorporate the completion methods. Approaches such as the Nearest Neighbor Method, Linear Regression, and Dense Layer Neural Networks were evaluated. The methodology included data preprocessing, exploratory statistical analysis, outlier treatment, normalization, and data separation. The results obtained were analyzed and compared using the RMSE (Root Mean Squared Error) metric. The findings showed that the ML-based technique achieved greater accuracy in imputing missing data compared to traditional methods. Dense layer neural networks demonstrated superior performance in capturing nonlinear patterns of climatic variables. The research suggests that integrating these methods into the existing API can significantly improve the quality of the climatic data provided, benefiting a wide range of applications.