A comparison of Deep Learning, Shallow Neural Network, and Principal Component Analysis based approaches to Thunderstorm Prediction

A deep learning neural network model was developed to predict thunderstorm occurrence within 400 km² South Texas domains with up to 15 h (+/-2h accuracy) lead time. The input features were selected or formulated primarily from select numerical weather prediction model output variables. Cloud-to-ground lightning data served as the target. The deep learning technique used is the stacked denoising autoencoder (SDAE) whereby hidden layers are pretrained in an unsupervised manner in order to create a higher order representation of the features. SDAE can be interpreted as a nonlinear approach to dimension reduction to select input features to a classifier by restricting the number of hidden layer neurons rather than the number of original input features. Logistic regression was applied to the SDAE output to predict thunderstorm occurrence. An iterative technique was used to determine the SDAE architecture corresponding to the greatest prediction performance; the model leading to the best prediction performance contained three latent variables. In addition to SDAE, Principal Component Analysis (PCA) was applied to the same dataset. It was found that SDAE generates more distinguishable latent features than PCA. Logistic regression was also applied to the PCA output to develop an alternative thunderstorm predictive model. The performance of both the SDAE and PCA based logistic regression models was compared for the foregoing lead times and 400 km² domain regions. The performance of the deep learning models is also compared to previously developed shallow neural network models and to forecasts from the National Weather Service Weather Forecast Office in Corpus Christi Texas using the same dataset. The superior performance of the deep neural network models over the shallow neural network models was evident. Deep learning and shallow neural network models yielded Peirce Skill Score (PSS) values between 0.70 and 0.78, and between 0.41 and 0.71, respectively. Corresponding Heidke Skill Score (HSS) values were between 0.49 and 0.58, and between 0.05 and 0.25, respectively.