Análise e previsão de dados meteorológicos com lacunas usando técnicas estatísticas e de aprendizado de máquina

Abstract

This study aims to fill gaps in precipitation and average temperature time series and generate 12-month forecasts from the completed series, using data from a meteorological station in Dois Vizinhos, Paraná, Brazil. Applied methods of linear interpolation and monthly historical average for smaller gaps. The forward-backward technique was employed, for the large gap, fitting exponential smoothing models, Box & Jenkins methodology and artificial neural networ to both segments. Subsequently, models using the same approaches were adjusted to the complete series to make forecasts for a 12-month period. The root mean square error between the predicted values and the actual data for the period was used as a basis for evaluation and comparing the models. The results indicated that the exponential smoothing model performed best for the precipitation series, while the neural network model was superior for the temperature series, indicating that neural network models have a better fit in series that present homogeneous seasonal behavior.