Probabilistic thunderstorm forecasts using statistical post-processing: Comparison of logistic regression and quantile regression forests and an investigation of physical predictors

Abstract

Probabilities of thunderstorm occurrence and conditional probabilities of lightning intensity over The Netherlands are forecast using statistical post-processing with predictors derived from the operational non-hydrostatic numerical weather prediction model Harmonie, at lead times up to 45 hours. Quantile regression forests (QRF) is compared with logistic regression (LR) for thunderstorm occurrence forecasts and with extended LR for lightning intensity forecasts. Using different sets of predictors that these statistical methods may select, it is demonstrated that pre-selection of predictors based on physical understanding and simultaneously exploiting QRF as machine learning tool can help improving statistical post-processing models. QRF is demonstrated to be beneficial for the predictions, with more skillful forecasts than LR for thunderstorm occurrence. Lightning intensity predictions are influenced by inhomogeneity of lightning detection datasets; despite inhomogeneity, skillful predictions can be made with both extended LR and QRF. The regional maximum of Modified Jefferson index and most unstable CAPE are found as best thunderstorm occurrence predictors and the regional minimum of Bradbury index and maximum of K-index emerge as best for lightning intensity. Neither most unstable CAPE nor microphysical predictors (graupel, snow) are essential for thunderstorm occurrence prediction.