Master's thesis defense by Peter Ukkonen – Niels Bohr Institute - University of Copenhagen

Forward this page to a friend Resize Print kalender-ikon Bookmark and Share

Niels Bohr Institute > Calendar > NBI Calendar 2017 > Master's thesis defens...

Master's thesis defense by Peter Ukkonen

Title: The Convective Parameterization Challenge

Atmospheric moist convection, and the precipitation it produces, is a very important phenomenon both in terms of societal impacts and the role it plays in the atmosphere. Yet, estimates of how convective precipitation will change in the future are associated with large uncertainties. Much of this uncertainty is due to convection being a parameterized process in climate models and many weather models. Many deficiencies have been identified in currently used convective parameterizations.
To circumvent these problems, non-linear machine learning methods could be used for modelling convective precipitation. This work investigates the use of artifical neural networks (ANN) for estimating hourly precipitation in a regional climate model. Fit-for-purpose hindcast simulations using a state-of-the-art RCM (Harmonie-Climate, HCLIM) are produced. These simulations span the summer months of 2011-2016 and are done on a 12 km grid using a Nordic domain. Two sets of neural networks are developed: one for predicting convective activity from various predictors calculated from RCM _elds. These ANN's are trained using lightning ash observations. The second set of ANN's predict hourly precipitation and are trained by an observational product which assimilates radar as well as rain-gauge data.
The results reveal some interesting features. One, thunderstorm activity can be predicted with relatively good skill using only a few simple inputs. This information is valuable to feed into a neural network to predict hourly precipitation, likely because it aids in the stratiform/convective separation.
Two, ANN's trained by minimizing mean squared error (MSE) on gridded data cannot produce a reasonable statistic of precipitation. While the ANN achieves a much lower MSE on an independent set of data compared to the RCM-predicted precipitation, it's practical use is unfeasible given that the model underestimates the intensity of moderate and heavy precipitation events by an order of magnitude. Much of the improvement in MSE was associated with bias removal, as HCLIM was in general too wet in the summer relative to observations.
The convective precipitation in HCLIM proved ine_ective as a predictor for observed rainfall, as it degraded overall MSE. The ANN, which used a few simple inputs such Lifted Index, was a more e_ective predictor. This suggests there are room for improvements in the current convective scheme. The results obtained here suggest ANN's may be useful for predicting convective initiation in numerical models.