Seasonal time-series imputation of gap missing algorithm (STIGMA)

Rangel-Heras, Eduardo ; Zuniga, Pavel ; Alanis, Alma Y. ; Hernandez-Vargas, Esteban A. ; Sanchez, Oscar D.

Kybernetika, Tome 59 (2023) no. 6, pp. 861-879

Résumé

This work presents a new approach for the imputation of missing data in weather time-series from a seasonal pattern; the seasonal time-series imputation of gap missing algorithm (STIGMA). The algorithm takes advantage from a seasonal pattern for the imputation of unknown data by averaging available data. We test the algorithm using data measured every $10$ minutes over a period of $365$ days during the year 2010; the variables include global irradiance, diffuse irradiance, ultraviolet irradiance, and temperature, arranged in a matrix of dimensions $52,560$ rows for data points over time and $4$ columns for weather variables. The particularity of this work is that the algorithm is well-suited for the imputation of values when the missing data are presented continuously and in seasonal patterns. The algorithm employs a date-time index to collect available data for the imputation of missing data, repeating the process until all missing values are calculated. The tests are performed by removing $5\%$, $10\%$, $15\%$, $20\%$, $25\%$, and $30\%$ of the available data, and the results are compared to autoregressive models. The proposed algorithm has been successfully tested with a maximum of $2,736$ contiguous missing values that account for $19$ consecutive days of a single month; this dataset is a portion of all the missing values when the time-series lacks $30\%$ of all data. The metrics to measure the performance of the algorithms are root-mean-square error (RMSE) and the coefficient of determination ($R^{2}$). The results indicate that the proposed algorithm outperforms autoregressive models while preserving the seasonal behavior of the time-series. The STIGMA is also tested with non-weather time-series of beer sales and number of air passengers per month, which also have a cyclical pattern, and the results show the precise imputation of data.

Détail

DOI : 10.14736/kyb-2023-6-0861

Classification : 62-04, 68Pxx
Keywords: contiguous missing values; seasonal patterns; time-series