Meteorological data can be brought into fire-growth models in a way that is both spatially and temporally accurate given the physical constraints of forecasted data; in doing so the predictive range of these models can be extended from the next few hours to days and even weeks. This was achieved by first examining the scales of fire operations and how weather forecasts are used at these scales. Approaches were then developed that would best use these forecasts in fire-growth models.

A suite of modelling approaches was developed to conduct predictions at three scales: short, medium and long range. These scales correspond to modelling fire growth on the scales of hours, days, and then weeks (or more). The first approach used a deterministic model using hourly meteorological data in the form of the model output statistics (MOS) based SCRIBE matrices provided by Environment Canada. The second approach introduced perturbations to the meteorological data brought into the growth model, accounting for a level of uncertainty to medium-range forecasts. The third approach included an original model to calculate the joint probabilities of fire spread and of fire survival from climatological data.

Validation studies were conducted to assess the skill of model predictions. Wood Buffalo National Park was chosen as the study site because of its natural fire regime Short-range predictions were tested for a pair of large fires in the park, which burned over 200 000 ha in June 2007. Results showed the model predictions had skill but that there were significant issues with using remotely-sensed hotspots to map the fire. For the medium-range, case studies were conducted showing that daily fire-growth predictions using maximum-minimum value weather forecasts over-predicted fire growth based on actual hourly weather observations (hindcasted) by 27%. Systematic-perturbation model runs best compensated for this with most fire growth falling within the predicted range of the models (52 out of 63 days). Long-range predictions were compared with distributions of fire perimeters predicted by repeated simulations using the hourly-based, deterministic fire-growth model. Close agreement between the long-range model and the deterministic model confirmed the probabilistic approach used by the long-range model.