Comparison between Continuous Integration and Targeted Initialization for Medium-Range Prediction of Severe Convective Storms

This study investigates the skill of a convection-allowing model (CAM) in forecasting severe convective storms (SCSs) using two different initialization approaches at lead days 4 and 8. Seven notable SCS events across the United States were simulated using the Weather Research and Forecasting Model, initialized with reforecast data from the Global Ensemble Forecast System, version 12. Continuous integration (CI) and targeted initialization (TI) methods were compared in an ensemble framework, employing two stochastic perturbation schemes to create three members. Forecast performance was evaluated through comparisons of accumulated precipitation, simulated radar reflectivity factor, and severe storm attributes against observed datasets, including the NCEP Stage IV quantitative precipitation estimates and GridRad reflectivity composites. Fractional skill scores revealed limited statistically significant differences between the two initialization techniques, although targeted initialization exhibited slight improvements for certain metrics, particularly at lead day 4. Furthermore, model simulations using targeted initialization substantially reduced computational wall-clock time, offering potential resource savings. These findings suggest that while both methods perform comparably, targeted initialization could be advantageous for operational CAM ensemble forecasting—particularly for medium- to extended-range lead times—where computational feasibility is often the biggest barrier. Significance Statement Advances in computational processing now permit the widespread operational use of convection-allowing models (CAMs) for short-range weather prediction. However, these simulations are still limited for medium- and extended-range forecasts, especially in ensemble frameworks, due to computational constraints. Targeted initialization, here referred to as a method to initialize a CAM using global model forecast data at a later forecast time, was compared to the traditional method of continuous integration. Results indicated that for large-scale severe-convective storm events driven by synoptic-scale features, there were limited statistically significant differences between the two initialization techniques, although targeted initialization exhibited slight improvements for certain metrics, particularly at lead day 4. Most importantly, targeted initialization presented an average compute time reduction of ∼61.5% and ∼73.7% for lead-day 4 and 8 forecasts, respectively, demonstrating that targeted initialization could serve as a valuable method for reducing the computational needs associated with operationalizing CAMs for medium-range forecasting applications.