Título: Regressão por Métodos Kernel na Previsão de Longo Prazo de Precipitação Pluviométrica Mensal
Autores: Allan Kelvin Sales, Guilherme de Alencar Barreto & Francisco de Assis de Souza Filho
Resumo: This article presents a systematic comparative analysis of sparse kernel adaptive filters (KAFs), specifically, KRLS-ALD, KRLS-T, SKRLS, QKLMS, and KNLMS, against traditional batch machine learning models (LSSVR, SVR- ν) for multi-step ahead monthly rainfall forecasting, using a univariate time series from Fortaleza, Brazil. As an original methodological contribution, we propose the IGNOS (Nonlinear Generalization Index for Relative Bias-Variance Tradeoff) metric to guide model selection by diagnosing overfitting and underfitting by quantifying the relationship between training and validation errors. The results demonstrate that while SVR- ν achieves a competitive test RMSE, KAFs exhibit significantly better computational efficiency and temporal robustness. In particular, the KRLS-ALD and KRLS-T algorithms better capture the non-stationary dynamics and extreme events that batch models tend to smooth out, establishing KAFs as a more practical and robust solution for hydrometeorological forecasting in challenging climates.
Palavras-chave: Long-term prediction; rainfall forecasting; Kernel methods; Sparse models; Machine learning.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191478
Artigo em PDF: CBIC_2025_paper1191478.pdf
Arquivo BibTeX:
CBIC_2025_1191478.bib