Baishemirov Z., Ospanova D., Amirgaliyev B., Mukhambetzhanov S.
Comparative analysis of physics-informed and conventional LSTM and RNN models for temperature forecasting using ERA5 reanalysis data // Journal of Physics and Chemistry of Sustainable Technologies in IT. — 2026. — Article No. 259. — DOI: 10.26577/jpcsit4120264.

Abstract: Climate change is one of the most serious modern problems affecting the Earth’s atmosphere, causing a range of global impacts. Forecasting climate dynamics is challenging due to the complexity and non-uniformity of climate data. While recurrent neural networks (RNNs) and long short-term memory (LSTM) models are widely used for time-series prediction, they often fail to satisfy physical constraints such as energy conservation and thermodynamic laws. This study develops physics-informed RNN and LSTM models and compares them with standard versions. The models are evaluated using ERA5 reanalysis temperature data for Astana, Almaty, and Shymkent. Results show that physics-informed models achieve lower RMSE in Almaty (3.52°C) and Shymkent (3.80°C), while standard models perform better in Astana (5.44–5.47°C). The results indicate that physics-informed approaches improve physical consistency, especially in stable climatic conditions, while conventional models may perform better in highly variable environments.

Link / DOI: https://doi.org/10.26577/jpcsit4120264