Astro-Cow-Net: An End-to-End Multimodal Deep Learning Model for Multi-Year Long-Lead El Niño Forecasting

Predicting the El Niño-Southern Oscillation (ENSO) over a long-lead horizon (12 to 24 months) is a notorious challenge due to chaotic atmospheric fluctuations. Traditional numerical weather models require intense supercomputing resources, yet frequently suffer from rapid error accumulation. This paper introduces Astro-Cow-Net, a purely data-driven, lightweight, end-to-end multimodal deep learning model designed specifically to forecast long-term El Niño and La Niña trajectories. By incorporating auxiliary orbital sequence inputs synchronized from NASA JPL ephemerides alongside standard historical ocean-atmosphere observations (Sea Surface Temperature and wind vectors), the sequential neural network establishes a resilient multi-source temporal mapping. Validated via a 5-fold cross-validation strategy, the model achieves a steady 24-month lead Mean Absolute Error (MAE) of 0.510, exploring multi-year forecasting capabilities on consumer-grade hardware.