Automation and Runtime Improvements for the Incident-based Scenario Evacuation (ISE) Collection of Models

Industry, academic, and government partners collaborated to produce significant improvements to the ISE suite of models. These models take NOAA GEFS ensemble forecasts and generate an ensemble of hurricane scenario tracks and associated wind and flooding hazard maps (the hazard models). They then use a progressive hedging algorithm to recommend a decision tree of evacuation orders optimized to minimize risk in an evacuation model (the evacuation scenarios model). This poster demonstrates the work plan to upgrade this suite of models from a research prototype to a working prototype, and the approaches for the hazard models and the evacuation scenarios model. The hazard models are a collection of three models. WRF (expand) generates hurricane track forecasts for each scenario. CREST (expand) takes WRFs precipitation forecasts and calculates the expected downstream flooding. Near coastlines, CREST hands off to ADCIRC (expand) which calculates the additional flooding due to storm surge. In the research prototype, these models are run at different sites across the country. File transfers and other overhead are a significant bottleneck. Further, not all hurricane scenarios are run concurrently due to resource constraints. The working prototype will run these scenarios concurrently from the same cloud-based HPC platform, greatly improving performance. We are completely rewriting the evacuation scenarios model. In the research prototype, it was a collection of scripts written in five different languages (Python, R, MatLab, Java, and Mathematica). These scripts operate on Microsoft Excel files and take 36 hours to run with significant manual interaction. The new version of the code is written entirely in Python, is fully automated, and is expected to run in approximately 2.5 hours. This is a 24x improvement.