ML for weather forecast
Title: ML for weather forecast
SNIC Project: Berzelius-2022-18
Project Type: LiU Berzelius
Principal Investigator: Ricardo Vinuesa <rvinuesa@mech.kth.se>
Affiliation: Kungliga Tekniska högskolan
Duration: 2022-02-01 – 2022-08-01
Classification: 20301
Homepage: https://www.vinuesalab.com/
Keywords:

Abstract

Today’s meteorological models use Finite Element Methods to partition the atmosphere into a 3D cube grid where every element is given some meteorological data based on measurements (windspeed, temperature, precipitation etc.). Then numerical models based on Navier-Stokes and chemical models are used to predict the future. AI is sometimes used today in the final stage to improve the result of these numerical predictions. I will build a Deep Neural Network that trains directly on the composite radar data (2D-maps). The network architectures I consider are: 1. 2D-Convolutional LSTM-network (easy to implement, hard to parallelize) 2. 3D-CNN including the temporal dimension (easy to implement, easier to parallelize) 3. A state-of-the-art Transformer model (harder to implement, easier to parallelize) I will investigate which method can give better predictions towards the development of new forecast tools within the scope of my thesis project. Iwill start with a thorough study of how the forecasts are done today, globally and at SMHI, as well as investigate what work has been done regarding AI and weather forecasting. An upside to this sort of solution for weather forecasting is that such a network would require less computing power than today’s models once the network is trained, which is a one-time cost. This means that the forecasts then can be based on a larger dataset than those used in conventional methods or be generated much quicker. The challenge will be that the dataset is huge with years of data, thus I will have to consider spatial down sampling, parallelization etc.