Evaluation of WRF for fine scale surface energy balance modeling in Phoenix

We utilize observations from our new flux tower situated within a residential neighborhood in Phoenix, Arizona, to evaluate the application of urbanized WRF for fine scale surface energy balance (SEB) modeling when incorporating 30-meter resolution NLCD urban land use data. Large seasonal variations in net radiation forcing were achieved by selecting synoptically-quiescent high-pressure periods in winter 2011 and pre-monsoon summer 2012. Examination of diurnal variation in SEB constituent fluxes, in conjunction with the diagnostic 2-meter temperature (T2), indicate that model performance cannot be assessed based solely on T2, an observed variable provided by many sensor networks, and as such, often employed within metrics of model performance. We discuss physical mechanisms of incongruency between model predictions and observations, with limitations constrained by: i) parameterizations of model physics, ii) parameterizations of input data, iii) model resolution, iv) observation resolution. Land use and land cover (LULC) were derived from combining 30-meter resolution NLCD 2006 urban classes with 1-kilometer resolution MODIS. We present simulations employing five nested domains to 333 meter horizontal resolution, with 40 vertical levels. Model configuration cases were comprised of the bulk scheme versus the single layer urban canopy model (UCM) for the default and two previously published parameterizations of Phoenix, and controlling for local (MYJ) and non-local (YSU) PBL schemes. We show that there are cases when, although the simulated temperature has good agreement with observations, the individual SEB terms can have considerable error. We found improvement when adapting UCM for Phoenix with reduced errors in the SEB components. Regarding T2, all cases we tested perform well during mid-day unstable convective periods for both seasons. Whereas, both UCM and PBL give rise to mixed benefit during other stability regimes, with YSU consistently closer to observations than MYJ except for when the bulk scheme was used with MYJ. We show that as model resolution is refined, the model remains stable, and most variables are statistically indistinguishable with scale, despite violating various assumptions.