Título: Norm-Cosine for Convex Combinations vs XGBoost on Climate Prediction of Belo Horizonte and Shenzhen
Autores: Luiz Fernando Ramos Lemos, Gabriel Victor de Lima, Matheus Brendon Francisco, Rafael de Magalhaes Dias Frinhani, Jose Arnaldo Barra Montevechi, Anderson Paulo de Paiva
Abstract: This work proposes the NC-CC (Norm-Cosine for Convex Combination) method as an approach for multi-horizon forecasting of climate time series, comparing it with the XGBoost model. NC-CC identifies the historical time point most similar to the current one using a hybrid metric based on Euclidean distance and cosine similarity between climate variation vectors and forecasts the next k days through a convex combination of the current and analogous variations. We use daily temperature, precipitation, and wind direction data from 2019 to 2024 for the cities of Belo Horizonte (Brazil) and Shenzhen (China). We evaluate NC-CC and XBoost in 1,000 random trials, with forecast horizons of 1, 7, 15, and 30 days, considering quality metrics such as Root Mean-Square Error (RMSE), Mean Absolute Percentage Error (MAPE), RMSE/σ and accuracy within absolute thresholds. The results show that NC-CC outperforms XGBoost in terms of scalability and in tracking abrupt variations, whereas XGBoost tends to minimize quadratic error, yielding smoother forecasts.
Palavras-chave: time series forecasting, climate prediction, XGBoost, NC-CC, convex combination, vector similarity.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1180185
Artigo em PDF: CBIC_2025_paper1180185.pdf
Arquivo BibTeX:
CBIC_2025_1180185.bib