Application and Comparison of Two Transformer-Based Deep Learning Models in Short-Term Precipitation Nowcasting

Against the background of intensifying global climate change, extreme precipitation events have become increasingly frequent. Improving the accuracy of short-term precipitation nowcasting is therefore essential for disaster prevention and mitigation. Traditional numerical weather prediction (NWP) approaches are constrained by computational latency and errors arising from physical parameterizations, making it difficult to satisfy real-time forecasting requirements at high spatiotemporal resolution. Using the SEVIR dataset, this study conducts a systematic comparison of two Transformer-based deep learning models—Earthformer and LLMDiff—for short-term extreme precipitation nowcasting. Model performance is evaluated using the Critical Success Index (CSI), Probability of Detection (POD), and Success Ratio (SUCR). Results indicate that, for 0–30 min lead times, Earthformer more efficiently captures both local and long-range spatiotemporal dependencies via its Cuboid Attention mechanism and shows a slight advantage for low-intensity precipitation. As the lead time extends to 60 min, LLMDiff demonstrates stronger longer-horizon skill due to its diffusion-based probabilistic modeling and a frozen large language model (LLM) module, which enhance the representation of uncertainty and longer-term evolution of precipitation systems. However, LLMDiff tends to produce a higher false-alarm rate. Overall, Earthformer is better suited for rapid early warning of light precipitation, whereas LLMDiff is more appropriate for high-accuracy nowcasting of heavy precipitation, offering useful insights for intelligent forecasting of extreme weather.