Ensemble using different Planetary Boundary Layer schemes in WRF model for wind speed and direction prediction over Apulia region

The Weather Research and Forecasting mesoscale model (WRF) was used to simulate hourly 10mwind speed and direction over the city of Taranto, Apulia region (south-eastern Italy). This area is characterizedby a large industrial complex including the largest European steel plant and is subject to a Regional Air QualityRecovery Plan. This plan constrains industries in the area to reduce by 10% the mean daily emissions by diffuseand point sources during specific meteorological conditions named wind days. According to the Recovery Plan,the Regional Environmental Agency ARPA-PUGLIA is responsible for forecasting these specific meteorologicalconditions with 72 h in advance and possibly issue the early warning.In particular, an accurate wind simulation is required. Unfortunately, numerical weather prediction modelssuffer from errors, especially for what concerns near-surface fields. These errors depend primarily on uncertaintiesin the initial and boundary conditions provided by global models and secondly on the model formulation, inparticular the physical parametrizations used to represent processes such as turbulence, radiation exchange, cumulusand microphysics. In our work, we tried to compensate for the latter limitation by using different PlanetaryBoundary Layer (PBL) parameterization schemes. Five combinations of PBL and Surface Layer (SL) schemeswere considered.Simulations are implemented in a real-time configuration since our intention is to analyze the same configurationimplemented by ARPA-PUGLIA for operational runs; the validation is focused over a time range extendingfrom 49 to 72 h with hourly time resolution. The assessment of the performance was computed by comparing theWRF model output with ground data measured at a weather monitoring station in Taranto, near the steel plant.After the analysis of the simulations performed with different PBL schemes, both simple (e.g. average) and morecomplex post-processing methods (e.g. weighted average, linear and nonlinear regression, and artificial neuralnetwork) are adopted to improve the performances with respect to the output of each single setup. The neuralnetwork approach comes out as the most promising method.