LPS Neural Operator (LPSNO): A Novel Deep Learning Framework to Predict
the Indian Monsoon Low-Pressure Systems
Abstract
The synoptic scale variability of the Indian summer monsoon (ISM) is
contributed by the weak cyclonic vortices known as low-pressure systems
(LPSs). LPSs are the primary mechanism by which central Indian plains
receive rainfall. Traditionally, synoptic variability is considered to
have a low predictability. In the present study, we developed a
framework, namely, LPS Neural Operator (LPSNO), using the neural
operator-based deep learning to predict the spatial structure of daily
mean sea level pressure anomalies over the Bay of Bengal at a resolution
of 1°x1°. The proposed neural operator extends the Fourier neural
operator framework by employing convolutional LSTMs in the operator
backbone. Further, the mean sea level pressure is reconstructed using
the predicted anomaly and the climatology, which is then used to track
the LPSs using a Lagrangian tracking algorithm. The median pattern
correlation between the predicted and actual mean sea-level pressure
anomalies over the BoB is about 88%, 60%, and 50% for 24, 48, and
72-hour forecasts, respectively. The proposed model improves the
accuracy of predictions compared with the earlier ConvLSTM models. The
pattern correlation between the observed and predicted synoptic activity
index (SAI) is 0.94, 0.9, and 0.87 for 1, 2, and 3-day ahead
predictions, respectively. A well-trained model of LPSNO takes only
~3.2 s to generate a one-day forecast on a single GPU
node of Nvidia V100, which is computationally extremely cheap compared
to the conventional numerical weather prediction models. The proposed
LPSNO can advance operational weather forecasting substantially.