Resumen

ARIAS-MUNOZ, Ana Cristina; COB-GARCIA, Susana  y  CALVO-VALVERDE, Luis-Alexander. Comparative analysis of traditional methods and a deep learning approach for multivariate imputation of missing values in the meteorological field. Tecnología en Marcha [online]. 2024, vol.37, n.3, pp.33-47. ISSN 0379-3982.  http://dx.doi.org/10.18845/tm.v37i2.6746.

Climate observations are the groundwork for several real-world applications such as weather forecasting, climate change monitoring and environmental impact assessments. However, the data is mostly measured and recorded by external devices exposed to numerous variables, causatives of malfunctions and, therefore, missing values. Nowadays, data imputation in the time series field has been researched in depth and a wide variety of methods have been proposed, where traditional classification and regression algorithms predominate, even though there are also deep learning approaches that manage to capture temporal relationships between observations. In this article, a comparative analysis between a classification imputation algorithm, a regression imputation algorithm, and a deep learning imputation model is made: MissForest algorithm, based on random trees; Expectation Maximization with Bootstrap (EMB), the maximum likelihood estimation algorithm; and a proposed deep learning model, based on the Long-Short Term Memory (LSTM) architecture. Data from the Costa Rica meteorological field were used, which consist of multivariate data coming from several weather stations in the same geographical area.

Palabras clave : Data imputation; EMB; MissForest; LSTM; time series.

        · resumen en Español     · texto en Español     · Español ( pdf )