Abstract This paper proposes a corrective algorithm for improving the accuracy of global horizontal irradiation (GHI) forecasts obtained from the numerical weather prediction (NWP) model of the European Centre for Medium-range Weather Forecast (ECMWF). Firstly, GHI forecasts were compared with experimental values from two ground-based stations located at the south of Portugal (Evora and Sines), and the influence of Sun-Earth geometry and atmospheric variables on the differences between predictions and measurements was analysed in order to identify the most relevant parameters. These differences are shown to be correlated mainly with the clearness index, solar zenith angle, mean air temperature, relative air humidity and total water column. Since the ECMWF model directly or indirectly provides all these variables, it is possible to estimate the bias of the predicted GHI as a function of forecast data taking as reference the measurements, which means that an algorithm based on correlations of such parameters can be used to correct forecasts in an operational time horizon. With that goal, an artificial neural network (ANN) based algorithm was developed in this work to improve GHI predictions, including also as input the global solar irradiation predicted by a reference clear sky model. The internal structure of the ANN was optimised, and a spatial and temporal downscaling procedure was also developed to obtain half-hour irradiation values. This algorithm was tested against the original ECMWF forecasts and a persistence model for four locations with different orography and climate as well as for various sky conditions (overcast, partly cloudy and clear sky), showing that it successfully improves the model predictions. Higher values of a Global Performance Index (GPI) based on seven statistical indicators and Forecast Score (FS) were found for the algorithm simulations, e.g. respectively 1.066 and 0.348 when considering all the locations and cloud cover conditions, while for the original ECMWF predictions the GPI is −1.874 and the FS is 0.282. This algorithm is useful if integrated into energy management tools of solar energy systems, namely low/medium temperature solar thermal and photovoltaic systems.