Site Adaptation for Global Horizontal Irradiance in Argentina Using a Multilayer Perceptron
Abstract
Accurately quantifying the spatial and temporal availability of the solar resource is a critical step before the sizing of large solar power plants can be optimized and their bankability established. To remove the inherent biases in time series of satellite-based irradiance estimates, short-term local measurements are used for calibration purposes along with a site adaptation (SA) algorithm. Here, a machine-learning algorithm based on Multilayer Perceptron (MLP) is used to improve estimates of global horizontal irradiance (GHI) at two sites in northwest Argentina. The SA algorithm is trained with one year of hourly data using six regressors, including GHI and cloud cover, to ultimately calibrate the CAMS-Rad satellite-derived estimates that are available in the public domain. The MLP-based adaptive function is then applied to several years of those GHI estimates and compared against ground measurements. Significant-to-large reductions in bias are found, most particularly at the high-elevation site of El Rosal, demonstrating the strength of the MLP approach in SA applications. The goal of this development is to help improve the viability and bankability of solar projects.
Keywords
GHI, site adaptation, MLP, machine learning, elevation, irradiance