Evaluation of Thunderstorm Predictors for Finland Using Reanalyses and Neural Networks

Abstract

AbstractThis work evaluates numerous thunderstorm predictors and investigates the use of artificial neural networks (ANNs) for identifying occurrences of thunderstorms in reanalysis data. Environmental conditions favorable for deep, moist convection are derived from 6-hourly ERA-Interim reanalyses, while thunderstorm occurrence in the following 6 h over Finland is derived from lightning location data. By taking advantage of the consistency and large sample size (14 summers) provided by the reanalysis, complex multivariate models can be trained for a robust estimation of convective weather events from model data. This and other methods are used to yield information on the most effective convective predictors in a multivariate setting, which can also benefit the forecasting community. The best ANN found uses 15 inputs and received a Heidke skill score (HSS) of 0.51 on an independent test sample. This is a substantial improvement over the best predictor when used alone, the most unstable lifted index (MULI) with HSS = 0.40, the multivariate model having fewer false alarms in particular. After MULI, the most important ANN input was relative humidity near 700 hPa. Dry air aloft was associated with significantly lower thunderstorm probability and flash density regardless of convective available potential energy (CAPE). Other important parameters for thunderstorm development were vertical velocity and low-level θe advection. Finally, the Peirce skill score indicates a clear meridional gradient in skill for categorical forecasts, with higher skill in northern Finland. This analysis suggests that the difference in skill is real and associated with a steeper thunderstorm probability curve in the north, but further studies are needed for a physical explanation.