Combining LSTM statistical analysis with dynamical models to investigate
Typhoon Mangkhut (2018)
Abstract
This study develops a long short-term memory mode (LSTM) neural network
algorithm for the prediction of tropical cyclone intensity (maximum wind
speed). This is achieved by combing the statistical tools of machine
learning and the dynamical earth system model of Typhoon Mangkhut
(2018). The Navy’s Coupled Ocean/Atmosphere Mesoscale Prediction System
is used to produce an ensemble of runs of Mangkhut to train the LSTM.
Being able to predict the behavior of tropical cyclones allows for a way
to analyze the complex dynamical inputs to investigate the driving
forces behind periods of rapid intensification, stagnation, and
weakening. The controlling parameters that produced the best prediction
are the mean inner surface heat flux, amplitude of the first order
asymmetry in the inner heat flux, the angle difference between the first
order asymmetries of vorticity output at 850mb and 500mb, and the
amplitude of first order asymmetry in the geopotential height at 850mb.