A hybrid approach for El Niño prediction based on Empirical Mode Decomposition and convolutional LSTM Encoder-Decoder. (April 2021)

• Record Type: Journal Article
• Authors: Wang, Si; Mu, Lin; Liu, Darong
• Is Part Of: Computers & geosciences. Volume 149(2021)
• Publication Date: 2021-04
• Subjects: El Niño -- Empirical mode decomposition -- Long short-term memory -- Oceanic Niño index -- Forecasting; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2021.104695
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: El Niño can affect climate patterns, causing extreme weather events, such as floods and droughts, around the world. Accurate forecasting of El Niño events allows preparation for El Niño-related disasters. However, the performance of current methods for predicting El Niño events one year in advance is not effective. This study proposes a hybrid approach to predicting the El Niño-related Oceanic Niño Index (ONI) and El Niño events with a lead time of 12 months. The proposed approach combines the convolutional Long Short-Term Memory (LSTM) Encoder-Decoder model with the Empirical Mode Decomposition (EMD) technique. The EMD technique can decompose time series data into a set of Intrinsic Mode Functions and a residue. Subsequently, the convolutional LSTM Encoder-Decoder model is employed to make an independent prediction for each component. Finally, the predicted data from each model can be reconstructed to obtain the forecasting results. The proposed approach is applied to the monthly ONI dataset for 1950–2019. The prediction model is trained and validated on the historical ONI values from 1950 to 2007 and forecasts El Niño events over a period of 12 years (2008–2019) with a lead time of 12 months. The results demonstrate that the proposed approach can successfully forecast that, for this period, 2009–2010, 2015–2016, and 2018–2019 are El Niño years. The performance of the proposed approach is then assessed by comparing it with the standalone convolutional LSTM Encoder-Decoder model, the LSTM-based models, and machine learning algorithms. The evaluated results indicate that the proposed approach outperforms these models in ONI predictions and El Niño event forecasts for a lead time of 12 months. Highlights: The approach based on deep learning can predict El Niño one year in advance. The proposed approach can achieve good performance for the ONI seq2seq prediction. The proposed approach can deal with the spring predictability barrier problem.