Censo: Cascade Time Series Forecasting Model for Enso Prediction

Abstract

Abstract-The prediction of the El Niño–Southern Oscillation (ENSO) index plays a crucial role in monitoring and early warning of global climate change. However, existing ConvLSTM-based recurrent prediction models suffer from rapid information decay and shallow dependency issues in long-term forecasting, leading to the accumulation of errors and decreased prediction accuracy. To address these challenges, this paper proposes a hierarchical recurrent prediction model based on a spatiotemporal LSTM network. Firstly, spatiotemporal memory units are incorporated into the recurrent prediction process, where the spatiotemporal state and forecasted output from the previous timestep are jointly fed as input to the next timestep, enhancing information transmission and mitigating information decay. Secondly, a cascade time series forecasting model (CENSO) is designed by integrating shallow-layer and deep-layer predictions, based on traditional autoregressive models, to achieve deeper fusion of temporal information between consecutive timesteps, thereby improving the stability and accuracy of long-term predictions. Experimental results demonstrate that the proposed model outperforms traditional recurrent models in multi-step ENSO forecasting tasks, showing higher accuracy and robustness. This approach offers new insights for complex predictions of oceanic spatiotemporal sequences and can be further applied to other fields related to climate change forecasting.

Keywords: ENSO prediction, spatiotemporal LSTM, recurrent prediction model, cascaded prediction, climate forecasting