Joint statistical clustering of spatio-temporal climatological data to predict wildfires and floods

Flooding and wildfires can create a compounding effect on damage to the power grid. Wildfires diminish the soil's ability to absorb water, leading to increased runoff during subsequent rainfall. To tackle this issue, we utilize a probabilistic model based on Markov Random Fields to identify both individual and joint clusters of climatological variables—such as wind speed, rainfall, and air temperature—using historical datasets from California. In this model, each local climatological variable is discretized according to multi-modal, data-driven probability distributions, and clustering incorporates weights for spatial and temporal coherence in addition to the observed values. Consequently, at each time step, the data from the selected geographical region is statistically aligned with one of a limited number of spatial clusters. This lower-order statistical representation enables us to effectively describe the complex patterns associated with multi-variable and heterogeneous weather events. The mechanisms underlying the correlations between wildfires and flooding and their impact on the power grid can be found from these clusters. This insight paves the way for developing predictive models that identify conditions leading to undesirable feedback between these two extreme events.