Performance Evaluation of a UAM-Specific High-Resolution Forecast-Data Assimilation Cycling System

This study aims to develop a high-resolution forecast–data assimilation cycling system to support Urban Air Mobility (UAM) operations. We implemented and evaluated a WRF-based 3DVAR–IAU (Incremental Analysis Update) cycling framework. Although 3DVAR is computationally efficient and suitable for high-frequency assimilation, directly incorporating the analysis into model integration can cause initial forecast discontinuities and spin-up issues. IAU mitigates these problems by gradually applying the analysis increment over the assimilation window.

The coupled WRF–WRFDA cycling procedure was automated to repeatedly perform 3DVAR analyses and subsequent forecasts using IAU. A preprocessing workflow was also established to process surface and vertical-profile observations, including LiDAR measurements, for data assimilation. To evaluate the performance of the system, we conducted two experiments: a CYCLE experiment (applying 3DVAR–IAU cycling) and a NOCYCLE experiment (a WRF-only free forecast without data assimilation). Forecast performance was assessed against observations using bias, root-mean-square error (RMSE), and correlation coefficients.

The results indicate that applying IAU reduces initial forecast discontinuities and leads to more stable early forecast behavior compared to NOCYCLE. Time–height cross-sections of wind speed error show that the CYCLE experiment generally produces smaller errors than NOCYCLE throughout the evaluation period. Consistently, the CYCLE experiment tends to yield lower RMSE and higher correlations relative to NOCYCLE for most vertical levels, indicating improved agreement with observations. Overall, these findings suggest that the proposed 3DVAR–IAU cycling approach can enhance the quality of assimilated initial conditions and contribute to continuous performance improvements in UAM-specific high-resolution prediction systems.

 

Key words: WRF, WRFDA, 3DVAR, IAU(Incremental Analysis Update), Cycling data assimilation