Machine Learning Forecast Products

The products are provided “as is” and are intended for research purposes only (disclaimer).

This page contains graphics designed to forecast the presence and strength of atmospheric rivers (ARs) and associated precipitation using innovative machine learning (ML) methods.

AR IVT Forecast Models:

Quantitative Precipitation Forecast (QPF) Models:

For more information, please contact Dr. Luca Delle Monache (Director of Research) or Dr. Agniv Sengupta (ML Team Lead).

Integrated Water Vapor Transport (IVT) with a Convolutional Neural Network (ARcnn)

Convolutional neural networks (CNNs) have been trained in a postprocessing framework for improving the National Center for Environmental Prediction (NCEP) Global Forecast System (GFS)’s integrated vapor transport (IVT) forecast in the eastern Pacific and western United States (Chapman et al. 2019, Geo. Res. Lett.). This method (GFS ARcnn) reduces errors in terms of RMSE at forecast leads from 3 hours to seven days (9–17% reduction) while increasing the correlation between observations and predictions (0.5–12% increase), and represents an approximately one‐to-two‐day lead time improvement.

Plot Description: Vertically Integrated Water Vapor Transport (IVT) with magnitude shaded in units of kg m-1 s-1 from the Global Forecast System (GFS; second row), Global Ensemble Forecast System mean (GEFS; third row), and the ARcnn (bottom row). For more information contact:……

Deterministic Precipitation Forecast with a U-Net Convolutional Neural Network

A U-Net convolutional neural network model has been developed based on a combination of classification and dual regression approach (Badrinath et al. 2023) to generate daily accumulated precipitation forecasts over the western U.S (0-4 day lead times). The model is trained using CW3E’s 34-year high-resolution deterministic West-WRF precipitation reforecasts and tested on four different water years conditioned on the state of the El Niño-Southern Oscillation (ENSO). Please refer to Badrinath et al. (2023) for the skill assessment (including skill for prediction of extreme events) associated with this forecast product.

Deep Learning-based Probabilistic Quantitative Precipitation Forecasts (PQPF)

A Deep Learning model, UNET, has been trained to postprocess CW3E’s 34-year deterministic West-WRF reforecast and generate probabilistic forecasts for 0-to-4-day daily accumulated precipitation (Hu et al. 2023, Mon. Wea. Rev.).

Plot Description: Top row: Ensemble mean 24-hour precipitation from the UNET colored in millimeters. Second row: Ensemble standard deviation of 24-hour precipitation shaded in mm from the UNET. Bottom five rows: Probability of 24-hour precipitation amount over various thresholds shaded in percentage. Probability is calculated based on the number of ensemble members predicting precipitation over the selected threshold.

Deep Learning of 200-member West-WRF Ensemble (200EnsDL) for Probabilistic Quantitative Precipitation Forecasts

A deep learning framework is introduced to postprocess CW3E’s 200-member West-WRF ensemble and to generate skillful, high-resolution, probabilistic forecasts for 0-to-6 days daily accumulated precipitation over the western United States (Ghazvinian et al. 2024, Mon. Wea. Rev.).

Plot Description: Top row: Ensemble mean 24-hour precipitation shaded in mm. Second row:Ensemble standard deviation of 24-hour precipitation shaded in mm. Bottom five rows: Probability of 24-hour precipitation exceeding various thresholds, e.g., >1 mm, >10 mm, >25 mm, >50 mm, and >100 mm, shaded in percentage (%).

Domain:  

Plot type: