Optimizing the Spatial-Temporal Extent of Environmental Factors in Forecasting El Niño and La Niña Using Recurrent Neural Network

Abstract

El Niño-Southern Oscillation (ENSO) is caused by periodic fluctuations in sea surface temperature and overlying air pressure across the Equatorial Pacific region. ENSO has a global impact on weather patterns and can cause severe weather events, such as heat waves, floods, and droughts, affecting regions far beyond the tropics. Therefore, forecasting ENSO with longer lead times is of great importance. This study utilizes Long Short-Term Memory (LSTM) network to predict ENSO events in the coming year based on environmental variables from previous years, including sea-surface temperature, sea level pressure, zonal wind, meridional wind, and zonal wind flux. These environmental variables are collected only inside certain spatial and temporal windows and used to forecast ENSO events. Furthermore, this study investigates how the size of these spatial and temporal windows influences the generalization accuracy of forecasting ENSO events. The size of spatial and temporal windows is optimized based on the generalization accuracy of the LSTM network in forecasting ENSO events. Our results indicated that the accuracy of the ENSO forecast is significantly sensitive to the extent of spatial and temporal windows. Specifically, increasing the temporal window size from one to nine years and the spatial window from 0 to 17.7 geographical degrees resulted in generalization accuracies, ranging from 40.1% to 83% in forecasting Central Pacific ENSO and 39.2% to 65% in forecasting Eastern Pacific ENSO.

Keywords

• climate anomalies
• machine learning
• spatial–temporal data mining
• deep learning
• ENSO events