Time Series Modeling Fire Radiative Power: Persistence or Machine Learning?

It has been previously shown that using machine learning, specifically random forests, are good tools for combining weather model output with satellite derived Fire Radiative Power (FRP). Their performance is comparable and often better than FRP persistence alone at forecasting next-hour FRP.

In this work, several machine learning (ML) models were trained with GFS weather data and the RAVE[1] FRP dataset inputs to forecast FRP 24 hours into the future. These models include random forest, gradient boosted, and Long Short-Term Memory (LSTM) structures and were then compared to the current FRP persistence modeling method, which was treated as the control. Different input combinations and model combinations were tested.

Both persistence and ML models were compared to observed hourly RAVE FRP to evaluate performance. Results found that especially for earlier forecast hours, ML models perform as well and often outperform the persistence method. Another finding is that the persistence modeled FRP follows a Guassian shape, and the error roughly follows that shape as well, but that the ML models are not confined to that shape and deviate from the forecasted FRP curve shape. This means that the ML models could provide more location and weather input unique FRP forecasts than the persistence method.

[1]https://doi.org/10.1016/j.rse.2022.113237