Wind Speed Time Series Imputation with a Bidirectional Gated Recurrent Unit (GRU) Model

Flores A.; Tito-Chura H.; Yana-Mamani V. · Lecture Notes in Networks and Systems (2022) · DOI: 10.1007/978-3-030-89880-9_34

Resumen

In this work, a novel bidirectional model based on the recurrent neural network known as Gated Recurrent Unit (GRU) is proposed for the imputation of not available (NA) values ​​in daily wind speed time series. The proposal model consists of two sequential GRU sub-models of 4-layers each, and for experimentation data from 3 years (2018–2020) is used, the first sub-model is trained with data from 2018 and the second sub-model with 2020 data, in both cases 2019 data is predicted, also, for second sub-model it's necessary the flipped 2020 data. Likewise, data augmentation is applied to improve the precision of the NA estimations. The results achieved show that the bidirectional proposal model achieves very good results, outperforming benchmark models such as Local Average of Nearest Neighbors (LANN), Autoregressive Integrated Moving Average (ARIMA) and Gated Recurrent Unit (GRU) without data augmentation. Likewise, comparing the results with other related works, it’s observed that proposal model surpasses most of them, making it an excellent alternative for wind speed time series imputation.

Instituciones: Universidad Nacional de Moquegua