Prediction of polar vortex intensity signal based on convolution smoothing and long short-term memory

Abstract

Polar vortex is an important weather system that affects the atmospheric circulation in the Northern Hemisphere and the climate change in the Arctic. The intensity variation of polar vortex is related to El Nino-Southern Oscillation (ENSO), Arctic Oscillation (AO) and many other climate phenomena. However, there are few researches on the prediction of polar vortex intensity change, our study analyzes and predicts the intensity variation of the Northern Hemisphere stratospheric polar vortex, and further uses convolution smoothing and depth learning methods to improve the accuracy of the prediction. The result shows that the long-short time memory network method’s prediction accuracy is not enough high. After the convolution smoothing of the time series of intensity signal, the prediction accuracy of neural network has been significantly improved. The average absolute error of the traditional long short-term memory network method is 18.29, while the average absolute error of the smoothed prediction intensity and the actual intensity is 13.77. In addition, the correlation between the predicted results and the real values is also as high as 0.9981.