Comparative Analysis of Medium-Range (72-hour) Temperature Time Series Forecasting Using Korea Meteorological Administration Surface Observation Data
Abstract
This study presents a comparative evaluation of medium-range (72-hour) air temperature forecasting performance among recent deep learning–based time series models—Neural Hierarchical Interpolation for Time Series Forecasting (NHITS), Temporal Fusion Transformer (TFT), Patch Time Series Transformer (PatchTST)—relative to a traditional statistical model (SARIMA) and a climatology-based baseline. The analysis is conducted using hourly surface observation data provided by the Korea Meteorological Administration, with a focus on a 72-hour multi-step forecasting task for the Seoul station.
Empirical results show that all three deep learning models substantially outperform the traditional baselines. Among them, the
TFT achieves the best overall performance, with a Mean Squared Error (MSE) of 4.39, a Root Mean Squared Error (RMSE) of 2.10, a Mean Absolute Error (MAE) of 1.65, and a coefficient of determination ( ) of 0.775. This corresponds to an approximate 54% reduction in RMSE compared with the SARIMA model. NHITS also demonstrates a significant improvement over the statistical baseline, achieving an RMSE of 2.61.
These results indicate that the Temporal Fusion Transformer architecture, which combines a Long Short-Term Memory (LSTM) based sequence encoder with multi-head attention, is particularly effective in capturing the nonlinear dynamics and combined seasonal and diurnal structures inherent in air temperature time series. This study provides quantitative evidence supporting the applicability of advanced deep learning models to short- to medium-range temperature forecasting based on domestic meteorological observation data. The findings have practical implications for applications requiring accurate temperature prediction, including smart grid energy demand management and weather-related disaster prevention. Future work should extend this framework to multivariate and multi-station settings, as well as to probabilistic forecasting approaches for explicit quantification of predictive uncertainty
References
- A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention Is All You Need,” in Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pp. 5998–6008, 2017.
- H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting,” arXiv preprint arXiv:2012.07436 [cs.LG], submitted Dec. 14, 2020, last revised Mar. 28, 2021. doi: 10.48550/arXiv.2012.07436.
- H. Wu, J. Xu, J. Wang, and M. Long, “Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting,” arXiv preprint arXiv:2106.13008 [cs.LG], submitted Jun. 24, 2021, last revised Jan. 7, 2022. doi: 10.48550/arXiv.2106.13008.
- Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (PatchTST),” arXiv preprint arXiv:2211.14730 [cs.LG], submitted Nov. 27, 2022, last revised Mar. 5, 2023. doi: 10.48550/arXiv.2211.14730.
- B. Lim, S. O. Arik, N. Loeff, and T. Pfister, “Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting,” International Journal of Forecasting, vol. 37, no. 4, pp. 1748–1764, 2021, doi: 10.1016/j.ijforecast.2021.03.012.
- Korea Meteorological Administration (KMA), “Surface Synoptic Observation Data (ASOS),” dataset.
- C. Challu, K. G. Olivares, B. N. Oreshkin, F. Garza, M. Mergenthaler-Canseco, and A. Dubrawski, “N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting,” arXiv preprint arXiv:2201.12886 [cs.LG], submitted Jan. 30, 2022, last revised Nov. 29, 2022 (v6), (accepted at AAAI-23). doi: 10.48550/arXiv.2201.12886.
- Yoon, J., & Jeon, M. (2017-11-24). Temperature Forecasting Model by using Deep Learning Technology based on LSTM. Conference of the Institute of Electronics and Information Engineers (IEIE), Incheon, Republic of Korea.
- Hong, S., Kim, J. H., Choi, D. S., & Baek, K. (2021). Development of Surface Weather Forecast Model by using LSTM Machine Learning Method. Atmosphere, 31(1), 73–83. https://doi.org/10.14191/ATMOS.2021.31.1.073
- Lim, J., Cho, D., Yoo, C., & Lee, Y. (2020, October 28). Improving next-day urban maximum temperature forecasts in a deep learning–based local forecasting model. Proceedings of the Korean Meteorological Society Conference, Seoul, Korea.
- Cho, D., Lim, J., & Jeong, S. (2023, November 1). Spatial refinement of summer short-term temperature forecasts by integrating deep learning and spatial interpolation methods. Proceedings of the Korean Meteorological Society Conference, Busan, Korea.
Keywords
Details
| Section | Articles |
| Issue | Vol. 2 No. 2 (2026): Volume 2 Issue 2 (Jun 2026) |
| Published | 2026-06-18 |
| Pages | 14-25 |
- Master of Engineering in AI Big Data, aSSIST (Seoul School of Integrated Sciences & Technologies)
- RESEARCH INTERESTS
- Large Language Models (LLMs)
- Transformers
- Time Series Analysis
- Visting Professor, Seoul School of Integrated Sciences & Technologies (aSSIST)
- Research Professor, The Institute for Industrial Policy Studies (IPS)
- Head of Process Management Team, Mirae Asset Securities
- Ph.D. in Business Administration, aSSIST University
- Research Interests: AI, Finance, Time Series, Business Strategy
License
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.