Melo, Leonimer Flávio deSilva, Ricardo Petri2026-09-242026-09-242024-07-12https://repositorio.uel.br/handle/123456789/20013Time series analysis proves highly useful and widely applicable, enabling reliable forecasting and informed decision-making across various fields. This thesis describes the fundamental importance of time series analysis and introduces advanced techniques aimed at enhancing the structure of time series data to improve predictive accuracy. Given the challenges associated with concept drifts—where the statistical properties of a signal change unpredictably—this study also emphasizes the need for innovative, unsupervised approaches to optimize drift detectors. The thesis presents two novel segmentation methods designed to refine data processing based on stationarity analysis, thus facilitating more accurate forecasts. Additionally, it develops a new method to enhance drift detectors using unsupervised approach, reducing the need for manual tuning and deep prior knowledge of the data. These contributions significantly advance the field of time series analysis, providing robust, adaptable, and efficient solutions for applications where precise predictions are vital. Our results demonstrate that the segmentation process enhances the prediction process for both Long Short-term Memory and Temporal Convolutional Networks. Nonetheless, our unsupervised tuning outperforms the default configuration in drift detectors in most evaluated cases.engRedes neurais (Computação)Análise de séries temporaisTeseEngenharias - Engenharia ElétricaEngenharias - Engenharia ElétricaTime series segmentationStationarity analysisTime series prediction improvementSize reduction in time seriesUnsupervised tuning of drift detectorsConcept driftsNeural networks (Computing)Time series analysis