Imputação multivariada de séries semporais meteorológicas do INMET com reanálise ERA5-land e self-attention

Abstract

Missing values in meteorological time series compromise climate analyses and predictive models. This work investigates a multivariate imputation pipeline for observational data from the Brazilian National Institute of Meteorology (INMET), using ERA5-Land as a source of auxiliary features. SAITS was adopted as the main model and compared with Median, LOCF, MissForest, XGBoost, Transformer and ImputeFormer. MCAR scenarios ranging from 10% to 80% artificial missingness were generated and evaluated using MAE, RMSE, and MASE. The results show that SAITS achieves the best performance under low missingness, while XGBoost obtains the lowest global errors from 40% to 80% missingness, indicating greater robustness under severe data degradation.