Spatiotemporal cross-attention hybrid network for weather forecasting in complex terrain

Area of Science:

• Meteorology
• Artificial Intelligence
• Geospatial Science

Background:

• Weather forecasting in complex terrain is challenging due to microscale topography and macroscale atmospheric dynamics.
• Existing models struggle to balance local spatiotemporal patterns with long-range dependencies.

Purpose of the Study:

• To develop a novel hybrid AI architecture for improved weather forecasting in complex terrains.
• To address limitations in capturing both local and global atmospheric patterns.

Main Methods:

• A hybrid ConvLSTM and Transformer architecture with a Spatiotemporal Cross-Attention (STCA) mechanism was developed.
• An STL-VMD preprocessing pipeline decomposed non-stationary meteorological signals.
• The model was validated across 33 stations in Guizhou over 51 years.

Main Results:

• The hybrid model significantly outperformed baseline LSTM models in temperature and precipitation forecasting.
• Achieved an 18.99% reduction in RMSE for temperature and a 27.74% reduction in RMSE for precipitation.
• Demonstrated a 54.08% improvement in [Formula: see text] for precipitation forecasting.

Conclusions:

• The novel hybrid architecture shows robustness and generalization capability for operational forecasting in complex terrain.
• The integration of ConvLSTM, Transformer, and STCA effectively captures spatiotemporal weather patterns.
• This approach offers a promising solution for enhancing weather prediction accuracy in challenging geographical areas.