MS-nowcasting: Operational Precipitation Nowcasting with Convolutional LSTMs at Microsoft Weather

  • Sylwester Klocek ,
  • Haiyu Dong ,
  • ,
  • Panashe Kanengoni ,
  • Najeeb Kazmi ,
  • Pete Luferenko ,
  • Zhongjian Lv ,
  • ,
  • Jonathan Weyn ,
  • Siqi Xiang

NeurIPS 2021 Workshop on Tackling Climate Change with Machine Learning |

Publication

We present the encoder-forecaster convolutional long short-term memory (LSTM) deep-learning model that powers Microsoft Weather’s operational precipitation nowcasting product. This model takes as input a sequence of weather radar mosaics and deterministically predicts future radar reflectivity at lead times up to 6 hours. By stacking a large input receptive field along the feature dimension and conditioning the model’s forecaster with predictions from the physics-based High Resolution Rapid Refresh (HRRR) model, we are able to outperform optical flow and HRRR baselines by 20-25% on multiple metrics averaged over all lead times.