Webinar: Weather forecast using Predictive Recurrent Neural Networks

Speaker
Dr. Hoang Viet Tran
Pacific Northwest National Laboratory

Abstract:

Numerical Weather Prediction (NWP) models provide reliable global forecasts but require high computational cost. Deep Learning (DL) emulators trained on NWP outputs offer cost-effective and reliable forecasts. However, current DL models oversimplify the temporal component leading to prediction errors accumulated quickly as the lead time increase.

We propose a DL architecture with more focus on weather variables’ dynamics through time. Our Predictive Recurrent Neural Network (PredRNN), previously used for modeling video dynamics, was trained on outputs of the European Centre for Medium-Range Weather Forecasts (ECMWF). Preliminary results show that the PredRNN outperformed the operational ECMWF and two latest DL emulators in forecasting wind speed, precipitation and sea surface pressure for hurricane Sandy. Our findings suggest that incorporating temporal dynamics in DL weather forecasting models can improve their accuracy.