Short-Term Quantitative Precipitation Forecacsting using Radar Data and Neural Networks
The short-term prediction of rainfall based on radar data is an important part of meteorology with uses that vary from flood prevention through traffic control to event planning. Modern systems allow for highly accurate approximation of the probability of rain within G hours, but not vet for the specific amount. The goal of the CIKM AnalytiCup 2017 is to improve existing systems to predict liquid precipitation quantities with 1 hour lead time by using exactly 1.5 hours of accumulated radar data.
Existing systems use computer vision algorithms to track convective cells and an empirical mathematical relation to convert the reflectivity values of these to rain rates, aside from complex physical simulations. Furthermore, they have more information available, such as wind direction and speed, temperature, atmospheric pressure, and topography. The first goal of this thesis is to solely use radar echo extrapolation data to achieve predictions with a RMSE lower than 14.G9 lnm/li. the second goal is to explore the performance of the statistical model used against a state-of-the-art Quantitative Precipitation Forecasting (QPF) system. Since the relationship between reflectivity values and the amount of rain is strongly non-linear, and given that artificial neural networks (ANN) are able to approximate any function in Mn, this thesis evaluated different reknown architectures of Convolutional Neural Networks as predictions models for the described problem. Although the set baseline was not reached, the comparison against a chosen industry-based QPF left a promising indicator for further research.