Use of Traditional and Neural-Network Based Post-Processing Techniques to Improve Fire-Indicator and Wildfire-Relevant Weather Variable Forecasts 2–6 Weeks Ahead

Statistical post-processing is a method to identify and reduce systematic biases and ensemble spread errors inherent in raw forecasts output from numerical weather prediction models such as the Global Ensemble Forecast System (GEFS). This presentation will first cover how we recently used traditional post-processing techniques to generate more skillful fire-indicator forecasts 8-14 days ahead for an experimental product running at the Climate Prediction Center. Those techniques used a parametric approach meaning that the data are fit to a representative probability distribution function. We will also show how we are currently applying nonparametric, artificial intelligence post-processing methods to improve Weeks 3-4 precipitation forecasts and how these methods will be applied in a similar manner to generate fire-indicator forecasts 2-6 weeks ahead. Specifically, these methods use artificial and convolutional neural networks to learn nonlinear relationships between various weather predictors and the resulting observation. These advanced post-processing techniques are critical to generating more reliable and skillful forecasts, especially at longer lead times. The post-processed forecasts can then be used to help identify regions over the contiguous United States that have probabilistic above- or below-normal wildfire potential in the weeks ahead.