Assessment of site-specific wind predictions by a hidden Markov model for multivariate circular-linear data

We consider a heavy industrial district close to the city of Taranto where winds from North-West quadrants and lack of precipitations are known to lead to a deterioration of urban air quality in terms of PM10 concentrations. In 2012, the Apulia Government adopted a Regional Air Quality Plan prescribing a reduction of industrial emissions by 10% every time such meteorological conditions are forecasted 72 hours in advance. In order to activate the appropriate safety measures, wind prediction is addressed by the Regional Environmental Protection Agency (ARPA Puglia) using the Weather Research and Forecasting (WRF) atmospheric simulation system. Here we investigate the ability of the WRF system to properly predict the local wind speed and direction allowing different performances for unknown weather regimes. Replicate observations of observed and WRF-predicted wind speed and direction at a relevant point location within the area of interest are jointly modeled as a multivariate 4-dimensional time series with a finite number of states (wind regimes) characterized by homogeneous distributional behavior. Observed and simulated wind data are made of two circular (direction) and two linear (speed) variables, then the 4-dimensional time series is jointly modeled by a mixture of projected-skew normal distributions with time-independent states, where the temporal evolution of the state membership follows a first order Markov process. Parameter estimates are obtained by a Bayesian MCMC-based method and results provide useful insights on wind regimes corresponding to different performances of WRF predictions.