Day ahead solar irradiance forecasting is essential to manage solar energy systems and assist decision-making. Available global and mesoscale Numerical Weather Prediction (NWP) models show limitations for regions with high complex terrain in solar irradiance forecasts, due to the usual coarse resolution of the models and the difficulties in reproducing the correct physics (in terms of cloud evolution) at the correct moment. This is the case of Reunion Island (21 degrees S, 55.5 degrees E) a 60-km wide tropical island located in the Southwest of the Indian Ocean basin. This work explores the use of NWP thermodynamic variables at synoptic scale as predictors of solar irradiance as they are commonly considered robust and well predicted by these models. Our analyses show that the diurnal cycle of solar irradiance on Reunion Island is sensitive to synoptic wind and relative humidity at 700 hPa (RH700). Moreover, this sensitivity varies depending on the site location on the Island with respect to synoptic wind direction. Hence, stations that are most exposed to synoptic wind, show greatest sensitivity. From these results, we have developed an empirical-statistical downscaling (ESD) method for day-ahead local-scale solar irradiance forecast and forecast reliability for the Reunion Island case, which searches for past analogous conditions of synoptic wind and RH700. Our study uses 12 years of ECMWF reanalysis data together with ground-based measurements at 21 stations on the island. Forecasting results for 4 of these stations plus an aggregate virtual station (derived as the mean of 11 coastal stations) show that the ESD method outperforms a climatology-based method in all cases. The contribution of this method for solar irradiance forecasting and its reliability are discussed. (C) 2015 Elsevier Ltd. All rights reserved.