Prediction of North Atlantic Oscillation Index Associated with the Sea Level Pressure Using DWT-LSTM and DWT-ConvLSTM Networks

Abstract EN

The North Atlantic Oscillation (NAO), which manifests as an irregular atmospheric fluctuation, has a profound effect on the global climate change.

The NAO index (NAOI) is the quantitative indicator that can reflect the intensity of the NAO events, and its traditional definition is the normalized sea level pressure (SLP) difference between Azores and Iceland.

From the variation tendency of the NAOI, we found that it is difficult to predict the NAO with the characteristics of variability and complexity.

As a data-driven approach, the deep neural network presents great potential in learning the mechanisms of climate forecasting.

In this paper, we adopt long short-term memory (LSTM) and ConvLSTM to predict the NAO from two aspects, NAOI and SLP, respectively.

In previous studies, LSTM has been regarded as a resultful method for time series prediction.

ConvLSTM can capture both the temporal and spatial interdependencies of the SLP field; then, the NAOI can be calculated from the SLP output.

In order to improve the prediction reliability, we utilize the discrete wavelet transform (DWT) as a preprocessing technique to decompose original data into different frequencies, considering the local time dependency.

It can effectively preserve the features of high-frequency data and forecast extreme events more accurately.

The proposed DWT-LSTM and DWT-ConvLSTM models are compared against multiple advanced models, such as LSTM, Holt-Winters, support vector regression (SVR), and gated recurrent unit (GRU).

The results indicate that both DWT-LSTM and DWT-ConvLSTM perform better, particularly at peak values.

As for the 31 NAO events from 2006 to 2015, our models achieve the lowest prediction error and the best stability.

Compared with the forecast products of CPC named Global Forecast System (GFS) and the ensemble forecasts (ENSM), our models are much closer to observation in multistep forecasting.