A deep learning model to predict thunderstorms within 400km2 south Texas domains
Abstract
High resolution predictions, both temporally and spatially, remain a challenge for the prediction of thunderstorms and related impacts such as lightning strikes. The goal of this work is to
improve and extend a machine learning method to predict thunderstorms at a 3km resolution with
lead times of up to 15 hours. A deep learning neural networks (DLNN) was developed to post process deterministic High-Resolution Rapid Refresh (HRRR) numerical weather prediction (NWP)
model output to develop DLNN thunderstorm prediction models (categorical and/or probabilistic
output) with performance exceeding that of the HRRR and other models currently available to National Weather Service (NWS) operational forecasters. Notwithstanding the discovery that shallow
neural network models can approximate any continuous function (provided the number of hidden
layer neurons is sufficient), studies have demonstrated that DLNN models based on representation
learning can perform superior to shallow models with respect to weather and air quality predictions. In particular, we use the method known as stacked autoencoder representation learning, yet
more specifically, greedy layer-wise unsupervised pretraining. The training domain is slightly
large specific area in Corpus Christi, Texas (CRP). The domain is separated into a grid of 13×22
equidistant points with a grid spacing of 20 km. These points serve as boundaries/centres for 286
20 × 20 km (400 km2 ) square regions. The strategy is that DLNN model train on whole boxes
and test on three most important boxes to evaluate the model. The target refers to the existence, or
non-existence, of thunderstorms (categorical). Cloud to Ground (CG) lightning was chosen as the
proxy for thunderstorm occurrence. Logistic regression was then applied to SDAE output to train
the predictive model. An iterative technique was used to determine the optimal SDAE architecture.
The performance of the optimized DLNN classifiers exceeded that of the corresponding shallow
neural network models developed by Collins and Tissot [12], a classifier via a combination of principal component analysis and logistic regression, and operational weather forecasters, based on the
same dataset.