Predição Conforme para Quantificação de Incertezas no Monitoramento de Estruturas Geotécnicas

Abstract

The monitoring of geotechnical structures generates a large amount of data, and machine learning techniques have been increasingly studied to predict and evaluate this data. However, machine learning-based models generally provide specific predictions without adequately addressing uncertainty. To better understand model uncertainty, the field of UQ has gained attention. In this work, we investigate the use of CP with the estimation of virtual sensors. Our study is divided into two stages: (1) the estimation of virtual sensors using machine learning techniques, and (2) exploring how CP can enhance the understanding of uncertainty in predictive models. We employ walk-forward validation with different time windows to maintain the temporal order of the data and adapt the multiple time series for CP application. The model’s performance is assessed using the Coefficient of Determination (R²), and the quality of the prediction intervals is evaluated through the PICP, PINAW, and CWC metrics. The results indicate that this approach warrants further investigation, as preliminary findings suggest that prediction methodologies can enhance the interpretation and utilization of predictive models.