LMcast: A pretrained language model guided long-term memory transformer for precipitation nowcasting

Deep learning models are effective in precipitation nowcasting. However, rainfall systems, as chaotic evolving systems, entail significant uncertainty. Previous studies have addressed precipitation nowcasting tasks by utilizing current image information from the perspective of overall motion trends or local rainfall details. As lead time increases, effective information decreases, making it challenging for existing methods to capture the long-term trends in rainfall systems. Recent studies have revealed that large language models demonstrate excellent retrieval and generation capabilities. Therefore, by leveraging the strengths of large language models and using similar historical rainfall processes as prior knowledge, the issue of insufficient effective information can be skillfully resolved. In this paper, we propose LMcast, a precipitation nowcasting model that utilizes a long-term memory recall approach guided by pre-trained language models. LMcast leverages the pre-existing linguistic knowledge of language models to recall historical future rainfall information from a codebook storing historical rainfall data. Moreover, we design a special fusion architecture to help LMcast combine recalled historical long-term memory with current input-generated short-term memory to produce the final prediction. Extensive experimental results on four publicly available radar datasets demonstrate the effectiveness and superiority of our proposed model compared to state-of-the-art techniques.