Traditional seasonal forecasting has predominantly relied on dynamical models; however, predictive skill deteriorates rapidly at longer lead times due to the chaotic nature of the atmosphere-ocean system. Recent advancements in Artificial Intelligence (AI) present new possibilities to overcome these limitations. This study proposes a novel seasonal forecasting framework for Sea Surface Temperature (SST). By employing Empirical Orthogonal Function (EOF) analysis, SST is decomposed into spatial modes and Principal Component (PC) time series. Each PC is subsequently predicted using a simple Neural Ordinary Differential Equations (Neural ODEs) model and reconstructed by combining it with the corresponding EOF modes. Comparative experiments demonstrate that the proposed Neural ODEs-based model exhibits more stable training and superior long-term forecasting potential compared to traditional RNN-based models (e.g., GRU), achieving an accuracy comparable to the dynamical North American Multi-Model Ensemble (NMME). Notably, in the Niño 3.4 region, the model maintains Anomaly Correlation Coefficients (ACC) exceeding 0.5 for lead times up to 11 months, confirming its capability for long-term prediction. To substantiate this predictability, the predictable time scales and inherent predictability within the PC data were analyzed. Multifractal Detrended Fluctuation Analysis (MF-DFA) was applied to diagnose the long-term memory characteristics of each PC. The analysis reveals that PC1 exhibits pink noise characteristics with a high predictable time scale, whereas higher-order modes display randomness akin to white noise, indicating intrinsic predictive limitations. Ultimately, these findings demonstrate that a simple Neural ODEs model performs competitively against dynamical models in long-term seasonal forecasting, and suggest that seasonal predictability inherently depends on mode-specific memory.
 

Neural ODEs,Seasonal Forecast,Memory characteristics,SST Forecast