Seasonal Precipitation Forecast based on a Machine Learning Approach

This CW3E forecast product predicts likely seasonal precipitation anomaly patterns for November-December-January over the western U.S. based on machine learning models and dynamical ensemble predictions. Four different machine learning models (Random Forest, XGBoost, LSTM, and neural networks) were pretrained based on learning relationships between remote ocean and atmospheric patterns and characteristic seasonal precipitation anomalies over the Western United States. Instead of training on observations, the machine learning models were here trained on a large climate model ensemble, which enabled more robust relationships to be established due to the very large sample size in the training phase (Gibson et al. 2021). Seasonal forecast skill is also enhanced by focusing on four characteristic large-scale precipitation patterns across the Western United States. Our historical skill assessment found that the probability of a correct forecast improves when the Machine learning models show stronger consensus for a particular pattern. As such, the ensemble agreement (or lack of) provides useful guidance on forecast uncertainty.

Top row: Four precipitation anomaly patterns identified in Gibson et al. 2021. Percentages represent the number of ensemble models predicting that pattern for November 2021 – January 2022 with the red outline and plot size representing the most predicted pattern.

Middle row: Three-month precipitation anomaly pattern forecasts from the four CW3E developed machine learning models valid for November 2021 – January 2022.

Bottom row: Three-month precipitation anomaly forecasts from five North American Multi-Model Ensemble (NMME) models valid for November 2021 – January 2022. The associated precipitation pattern for each forecast is noted in the bottom right of each image.

Possible Precipitation Patterns

CW3E Machine Learning Models

North American Multi-Model Ensemble