Modelos de preenchimento de falhas de dados pluviométricos diários usando aprendizado de máquina.

Abstract

The reconstruction of incomplete rainfall time series is crucial for both Climatology/Hydrology and Civil Engineering. Data quality and integrity directly impact the accuracy of weather forecasts and hydrological investigations, as well as the proper design of water infrastructure such as reservoirs, dams, and urban drainage systems, whose precision relies on the reliability of rainfall data. This study investigates the application of machine learning models, namely KNN Imputer and Iterative Imputer, for imputing missing rainfall data, comparing them to traditional methods. Datasets covering the periods from 2002 to 2019 (collected at the Itirapina-CRHEA station) and 1979 to 2023 (provided by the Brazilian National Institute of Meteorology (INMET) in São Carlos) were used. Three distinct approaches were employed: a global analysis considering the entire dataset, and separate analyses for the wet and dry seasons to evaluate the methods' performance under different seasonal regimes. To validate the models, percentages of the data (2.5%, 5.0%, 7.5%, and 10.0%) were removed, simulating missing values, and the imputed values were compared to the originals. A detailed comparison between the machine learning models and traditional techniques was also performed. The results demonstrate that the machine learning models outperformed conventional methods in terms of accuracy and maintaining data consistency, proving effective in improving data quality and increasing the reliability of predictions, and highlighting their potential for application in broader climate studies. For future research, the application of these methods in different geographical regions, biomes, and climate regimes is suggested, which will allow for the evaluation of the adaptability and efficiency of the proposed methods in diverse contexts, contributing to water resource management and long-term climate planning in Brazil.