Browsing by Author "Ukkonen, Peter"

While numerical weather forecasts have improved dramatically in recent decades, forecasting severe weather events remains a great challenge due to models being unable to resolve convection explicitly. Forecasters commonly utilize large-scale convective parameters derived from atmospheric soundings to assess whether the atmosphere has the potential to develop convective storms. These parameters are able to describe the environments in which thunderstorms occur but relate to actual thunderstorm events only probabilistically. Roine (2001) used atmospheric soundings and thunderstorm observations to assess which from a variety of stability indices were most successful in predicting thunderstorms in Finland, and found that Surface Lifted Index, CAPE and the Showalter index were most skillful based on the data set in question. This study aims to extend the assessment of thunderstorm predictors to atmospheric reanalyses, by utilising model pseudo-soundings. Reanalyses such as ERA-Interim use sophisticated data assimilation schemes to reconstruct past atmospheric conditions from historical observational data. In addition to a large sample size, this approach enables examining the use of other large-scale model parameters, which are hypothesized to be associated with convective initiation, as supplemental forecast parameters. Using lightning location data and ERA-Interim reanalysis fields for Finnish summers between 2002 and 2013, it is found that the Lifted Index (LI) based on the most unstable parcel in the lowest 300 hPa has the highest forecast skill among traditional stability indices. By combining this index with the dew point depression at 700 hPa and low-level vertical shear, its performance can be further slightly increased. Moreover, vertically integrated mass flux convergence between the surface and 500 hPa calculated from the ERA-I convergence seems to have high association with thunderstorm occurrence when used as a supplementary parameter. Finally, artificial neural networks (ANN) were developed for predicting thunderstorm occurrence, and their forecast skill compared to that of stability indices. The best ANN found, utilizing 11 parameters as input, clearly outperformed the best stability indices in a skill score test; achieving a True Skill Score of 0.69 compared to 0.61 with the most unstable Lifted Index. The results suggest that ANNs, due to their inherent nonlinearity, represent a promising tool for forecasting of deep, moist convection.