Speaker Deck

Deep Learning for Rain and Lightning Nowcasting @NIPS2016

by Valerio Maggio

Published December 10, 2016 in Research

We describe a deep learning framework for precipitation and lightning nowcasting, applied to weather echo radar and lightning data at regional scale in Trentino-Sudtirol, in the Italian Alps. Nowcasting, i.e. forecasts obtained by extrapolation for a period of 0 up to 6 hours ahead, is based on a Convolutional Long-Short Term Memory model (ConvLSTM) (Shi et al., 2015) and it is embedded in an operational context. The model is able to forecast reflectivity up to 75′ ahead at a spatial resolution of 1.65 km based on 5 frames of recent (25′) radar data. Further, the model can manage blocking effects due to orography. The framework has been also applied to 95.7K events collected from a collaborative lightning location network on the same region, with a predictive accuracy of CSI=0.585, comparable with state of the art machine learning based solutions.