[Paper Review] MS-nowcasting: Operational Precipitation Nowcasting with Convolutional LSTMs at Microsoft Weather

Sylwester Klocek, Haiyu Dong+8 more|arXiv (Cornell University)|Nov 18, 2021

Meteorological Phenomena and Simulations17 references17 citations

TL;DR

MS-nowcasting is a ConvLSTM-based deep learning model that improves operational precipitation nowcasting by integrating a large spatial receptive field (LV) and HRRR physics-based forecasts, achieving 20-25% better performance than optical flow and HRRR baselines across multiple metrics over 6-hour lead times, while running efficiently on a single GPU.

ABSTRACT

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.

Motivation & Objective

• To develop a high-accuracy, low-latency precipitation nowcasting system for operational use in Microsoft Weather.
• To overcome limitations of traditional NWP models like HRRR, which suffer from spin-up delays and limited skill beyond a few hours.
• To improve upon radar-only deep learning methods by incorporating long-range spatial context and physics-based prior knowledge.
• To enable real-time, high-refresh-rate forecasting with minimal hardware requirements, supporting early warning for extreme weather.
• To achieve state-of-the-art performance in short-term precipitation prediction using a computationally efficient architecture.

Proposed method

• The model uses a three-layer encoder-forecaster ConvLSTM architecture with input sequences of radar reflectivity from the MRMS dataset.
• A novel large viewport (LV) technique stacks 25 spatial segments of a 1280×1280 km radar field into a 25×256×256 feature tensor to increase spatial context.
• The forecaster is conditioned on HRRR radar reflectivity forecasts, providing physical consistency and improving prediction of precipitation formation and dissipation.
• The model employs a combination of MAE and MSE loss with B-MAE thresholds, Adam optimizer with stochastic weight averaging, and gradient clipping for stable training.
• Training and inference are accelerated using the DeepSpeed framework for multi-GPU efficiency and low-latency inference.
• The model is trained on 10 million 440-minute sequences from 2018 MRMS data, with 80 min input and 360 min target, and evaluated on 2019 test data.

Experimental results

Research questions

• RQ1 Can a deep learning model that integrates large spatial context and physics-based forecasts outperform radar-only and NWP baselines in precipitation nowcasting?
• RQ2 How does conditioning a ConvLSTM forecaster with HRRR output affect prediction accuracy and bias in short-term precipitation forecasts?
• RQ3 What is the impact of increasing the input receptive field beyond standard spatial viewports on nowcasting performance?
• RQ4 Can a high-accuracy nowcasting model be deployed operationally with minimal hardware, achieving sub-2-minute latency on 8 GPUs?
• RQ5 How do different loss functions and optimization strategies affect the smoothness and realism of generated radar reflectivity forecasts?

Key findings

• The MS-nowcasting model with both large viewport (LV) and HRRR conditioning achieved a mean absolute error (MAE) of 3.23 dBZ over the first 2 hours, outperforming all baselines including HRRR (5.54 dBZ) and optical flow (4.39 dBZ).
• The model with LV alone reduced MAE to 3.28 dBZ and improved F1 score to 0.808 at 12 dBZ, demonstrating the value of extended spatial context.
• Adding HRRR conditioning improved MAE to 3.55 dBZ but slightly reduced F1 score (0.780), indicating a trade-off between bias and threshold detection.
• The full model (LV + HRRR) achieved the best overall performance with MAE of 3.23 dBZ, F1 of 0.813, MS-SSIM of 0.6721, and PSNR of 24.03, outperforming all baselines across all metrics.
• The model achieved end-to-end forecast generation in ~6 minutes from radar data availability, with inference latency under 2 minutes on 8 GPUs.
• Despite high performance, the model’s predictions were overly smooth due to MAE+MSE loss, suggesting potential for improvement with GAN-based loss functions.