Multi-step Temperature Forecasting with an Attention-based Gated Recurrent Unit Model

Kara V., Üstündağ Şiray G.

12th International Conference on Advances in Statistics, Barcelona, İspanya, 10 - 12 Nisan 2026, ss.1-2, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: Barcelona
  • Basıldığı Ülke: İspanya
  • Sayfa Sayıları: ss.1-2
  • Çukurova Üniversitesi Adresli: Evet

Özet

Temperature forecasting is a crucial problem for climate analysis, energy planning, and environmental decision support systems due to the nonlinear structure and long-term dependencies inherent in time series data [1]. In recent years, deep learning–based models have been increasingly adopted in time series forecasting to capture complex temporal relationships. Within this context, attention mechanisms have become increasingly important in allowing models to focus on the most informative temporal patterns, thereby enhancing forecasting performance [2][3]. Accordingly, this study proposes a deep learning framework based on a gated recurrent unit architecture enhanced with an attention mechanism for temperature forecasting using multivariate meteorological data [4].

The data set consists of daily temperature, wind speed, humidity, and surface pressure variables obtained from the NASA POWER database for a specific geographical location over the period from 2020 to 2024 [5]. To assess the effectiveness of the proposed approach, a comparative analysis is conducted with commonly utilized deep learning models reported in the literature, namely long short-term memory and gated recurrent unit architectures.

The comparative analysis is conducted under a multi-step forecasting framework considering prediction horizons of 7, 14, and 30 days. Model accuracy is evaluated using two metrics, which are Root Mean Squared Error and Mean Absolute Error. The empirical results demonstrate that the attention-based gated recurrent unit model achieves lower error rates across all forecasting horizons, exhibiting superior predictive performance compared to the baseline models.

In conclusion, this study indicates that the attention-based gated recurrent unit model provides an effective and reliable approach for temperature forecasting using meteorological time series, outperforming commonly used basic deep learning models.