Abstract

The integration of photovoltaic plants into the energy matrix is increasing around the world. In some of these matrices, the electricity grid operators need to know when the PV plant can input energy into the system and how much energy there will be. These types of technologies are directly affected by the variability of the solar resource, leading to possible destabilization of the electricity grid. Estimating realtime PV production is essential for improving the performance and operation of such facilities. Various research studies have used estimation techniques based on self-learning algorithms or statistical models. This work develops a methodology that allows us to evaluate the energy viability before installing a PV plant. The novelty of this work is that it evaluates three different statistical techniques - an Artificial Neural Network, a Support Vector Machine and a Multiple Linear Regression e in estimating the production in three PV plants located in Almería (Spain), Antofagasta and San Pedro de Atacama (Chile). This has been achieved by developing local models, where atmospheric variables are introduced into the different techniques to determine the PV production. The normalized root-mean square error statistical index presented values close to 3% in all cases. To facilitate the extrapolation of the models, a final global model was provided. This was trained with all the PV-plant data. It presented closer nRMSE values than those obtained from the local models, and the SVM results were slightly better. Consequently, we have a created a tool that can be used by companies, and the photovoltaic sector in general, to correctly size a plant and to estimate the final yield. This is achieved by accounting for the overall losses that are incurred, using the Performance Ratio (PR), thus providing a real study that serves as the economic basis for the investment and its benefits.