Co-Designing Short-Term Forecasts of Weather-Related Fire Brigade Operations

Convective events, windstorms, and heavy precipitation frequently trigger fire brigade operations, including responses to fallen trees, flooded basements, and infrastructure damage. Such events often lead to sharp increases in emergency call volumes, potentially overwhelming control and dispatch centers. Anticipating operation volumes through impact forecasts could support operational decision-making, for example by enabling timely staffing adjustments to manage peak demand.
Within the WEXICOM project, we develop forecasts of weather-related fire brigade operations through a co-design process involving five German control centers. This process includes expert interviews, iterative model development, prototyping, and user testing. In an initial co-design cycle, we developed a simple Poisson regression model and implemented it in a real-time prototype application. The model combines convective nowcast data with building coverage as predictor variables to estimate operation volumes. User testing yielded positive feedback; however, model validation revealed that this simple approach tends to underestimate operation counts.
In a second co-design cycle, we investigate how additional meteorological and exposure data improve the predictive skill of statistical models for forecasting weather-related fire brigade operations at lead times of 1–6 hours. Meteorological inputs include convective nowcasts, ensemble forecasts of wind and precipitation, radar-based precipitation estimates, and station observations. Exposure variables comprise building coverage and counts, road length, land use, and topography. Results indicate that predictor importance varies with lead time: nowcasts dominate at short lead times (1 hour), whereas ensemble forecasts provide the greatest skill at longer lead times. Stepwise regression is applied to select relevant predictors from a large candidate set. The combined models reduce the logarithmic mean absolute error by more than 30%.
We further compare Poisson, negative binomial, and hurdle models to account for overdispersion and excess zeros.  Overall, the hurdle negative binomial models provides the most appropriate representation of the data distribution.
To complete the co-design process, we will evaluate the impact of operation forecasts on the risk perception and decision-making of emergency managers in an experimental setting.