Large language models for the marine environmental variables prediction

Abstract

Monitoring and predicting marine environmental variables are important for safeguarding livelihoods and the economy. Large language models (LLMs) have shown great potential in time series prediction because of their strong computational capabilities, and the application of LLMs to the prediction of marine environmental variables is an emerging area of research. However, LLM-based approaches often exhibit oscillations in prediction outputs and large deviations from observed values. To address these issues, we propose TimeLLM-BERT, a hybrid three-stage model based on feature extraction, autoregressive prediction, and error correction. The model incorporates a structured prompt module, a trend enhancement algorithm, and a residual-fitting optimization strategy, which significantly enhance prediction accuracy. To systematically evaluate the performance of the model, comparative experiments were conducted against LSTM, BiTCN, NBEATSx, iTransformer, NHITS, and Time-LLM models using four key variables: significant wave height (SWH), sea surface temperature (SST), temperature at 2 m above the sea surface (T2M), and wind field (WF). The results show that the performance of the model is significantly better than existing models, and the mean absolute error for SWH prediction is reduced by 24.7%. It also achieves stable performance in SST prediction and strong consistency in WF prediction compared with the existing models. Robustness and universality tests show that the error evaluation indicators exhibit low variation, demonstrating strong stability and generalization ability. In summary, TimeLLM-BERT offers significant improvements in accuracy and stability for predicting marine environmental variables, providing a new framework for modeling complex time series data.