Análise e previsão de incêndios florestais para região de Lages e Xanxerê - Santa Catarina: uma abordagem explorando modelos híbridos de previsão

• Ano de defesa: 2025
• Autor(a) principal: Rocha, Vinicius Lovatel
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Pato Branco; Brasil; Programa de Pós-Graduação em Engenharia de Produção e Sistemas - Multicampi Pb e Ct; UTFPR
• Palavras-chave em Português: Incêndios florestais; Incêndios florestais - Previsão; Aprendizado do computador; Corpo de bombeiros - Santa Catarina; Forest fires; Forest fire forecasting; Machine learning; Fire departments - Santa Catarina; CNPQ::ENGENHARIAS::ENGENHARIA DE PRODUCAO; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/38592

Wildfires annually impact vast areas of the state of Santa Catarina, causing significant damage to the environment, economy, and both public and private property. This study aims to forecast wildfire occurrences in the regions of Lages and Xanxerê, which account for a high number of recorded incidents in the state. Based on operational data provided by the Military Fire Department of Santa Catarina (CBMSC), the research seeks to support more efficient mana- gement of human and material resources, enabling optimized personnel allocation during pe- riods of increased risk.The study is structured in two stages. The first involves conducting a systematic literature review to identify the main forecasting methods applied, the results ob- tained, observed limitations, and research gaps in the field of wildfire prediction. The second stage comprises an empirical study focused on the analysis and forecasting of wildfire out- breaks in the Lages and Xanxerê regions, employing both univariate and multivariate models, combined with the time series decomposition technique known as Variational Mode Decom- position (VMD).Model performance is assessed using evaluation metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (sMAPE), and Standard Deviation (SD). The analysis considers climatic variables such as precipitation, temperature, relative humidity, wind speed, and the occurrence of the La Niña phenomenon.Multivariate models showed better performance when applied without VMD. In Lages, the best result was achieved for forecasting horizon H = 2 using transformed data, with the ELM model proving most effective (RMSE = 3.1084; MAE = 2.1924; sMAPE = 61.96%; SD = 2.8475). In Xanxerê, also for H = 2, the XGBoost model without transformation was the most accurate (RMSE = 4.4526; MAE = 2.9776; sMAPE = 73.43%; SD = 4.5680).On the other hand, the univariate modeling approach, combining VMD with regression algorithms, proved more robust for monthly forecasting (H = 1). In Lages, the best performance was ob- tained using square root–transformed data, with the VMD-LASSO model standing out (RMSE= 1.2074; MAE = 0.9675; sMAPE = 58.90%; SD = 1.2401). In Xanxerê, the best results were also obtained with original (non-transformed) data, again highlighting the VMD-LASSO model (RMSE = 1.5248; MAE = 1.3597; sMAPE = 67.79%; SD = 1.4683).This study is expected to serve as a decision-support tool for CBMSC, contributing to operational planning and person- nel mobilization during critical periods, thereby strengthening wildfire prevention and response strategies.