Forecasting Daily Fire Radiative Energy Using Data Driven Methods and Machine Learning Techniques

Abstract

Increasing impacts of wildfires on Western US air quality highlights the need for forecasts of smoke emissions based on dynamic modeled wildfires. This work utilizes knowledge of weather, fuels, topography, and firefighting, combined with machine learning and other statistical methods, to generate 1- and 2-day forecasts of fire radiative energy (FRE). The models are trained on data covering 2019 and 2021 and evaluated on data for 2020. For the 1-day (2-day) forecasts, the random forest model shows the most skill, explaining 48% (25%) of the variance in observed daily FRE when trained on all available predictors compared to the 2% (<0%) of variance explained by persistence for the extreme fire year of 2020. The random forest model also shows improved skill in forecasting day-to-day increases and decreases in FRE, with 28% (39%) of observed increase (decrease) days predicted, and increase (decrease) days are identified with 62% (60%) accuracy. Error in the random forest increases with FRE, and the random forest tends toward persistence under severe fire weather. Sensitivity analysis shows that near-surface weather and the latest observed FRE contribute the most to the skill of the model. When the random forest model was trained on subsets of the training data produced by agencies (e.g., the Canadian or US Forest Services), comparable if not better performance was achieved (1-day R2 = 0.39–0.48, 2-day R2 = 0.13–0.34). FRE is used to compute emissions, so these results demonstrate potential for improved fire emissions forecasts for air quality models.