Statcasting: A machine learning based methodology for post-processing ensemble predictions of direct normal solar irradiance

Abstract

The inherent variability of Direct Normal Irradiance (DNI) exerts substantial influence one hand, on the grid integration of electric power generation by Concentrating Solar Technologies (CST) and the large-scale integration of these technologies into current energy supply structures; on the other side it presents high influence on activities related to the management and operation of concentrating solar power (CSP) plants whose technologies show a nonlinear response to DNI governed by various thermal inertias due to their complex response characteristics. Therefore, useful DNI forecast is of utmost importance for the management and operation of CSP plants, the power generation control by means of thermal energy storage (if available) and ultimately the efficient management of energy markets. It is usual the implementation of a prediction scheme that consists on a suitable post-process applied to deterministic weather forecast with the final goal of generate Solar Irradiance forecasts over a concrete emplacement. However, every numerical model prediction has an intrinsic uncertainty mainly due to the initial conditions and assumptions in model formulation. This fact has produced the development of different Ensemble Prediction Systems (EPS) which are designed to take into account this uncertainty providing a discrete sample of probability density function (PDF) of the meteorological parameters instead a single value. In this work the first approach of the Statcasting procedure is presented. It tries to extend the classical deterministic approach to the novelty paradigm of probabilistic forecast. This approach will allow the availability of uncertainty information of DNI and, therefore, a greater level of information in the decision making process inherent in the operation and management activities of the CSP plants.