Multi-Step ENSO Forecasting Using Kolmogorove Arnold Networks

ENSO (El Niño-Southern Oscillation) is a global climatic phenomenon that affects weather, agriculture, water supplies, and economies worldwide. Effective planning and mitigation need accurate ENSO forecasts. This study introduces the Kolmogorov-Arnold Network for Time Series Forecasting (KAN-TSF), which improves multi-step ENSO forecasting by combining KANs with LSTM networks’ gating mechanism. We present a multi-step time series forecasting model to capture detailed temporal patterns and manage non-linearity of time series data. The model is tested using ENSO dataset for ONI patterns from January 1950 to June 2024. KAN-TSF was compared to CNN, RNN, LSTM, and CNN-LSTM. The findings shows that KAN-TSF can successfully forecast ENSO trends over numerous time steps with the greatest R-squared value (0.902), lowest Loss (0.003), MAE (0.043), MAPE (12.220), and RMSE (0.056) among the other models, thereby demonstrating its superior capacity to accurately predict ONI patterns over multiple time steps.