Data-driven modeling for effective integration of wind power

Abstract

Stochastic methods to explicitly model the uncertainty of wind power output are essentialfor the effective integration of this renewable energy source into the power supply. Recent years have seen several advances in forecasting models for wind power output, though these models typically focused on a single wind farm that has already been built. A result of this dissertation is a novel forecasting framework that builds probabilistic forecasts for existing farms. This spatio-temporal framework exploits their asymmetric association in time and space via an ARIMA (Auto-Regressive Integrated Moving Average) -DVINE (Drawable Vine) copula model hybridized with support vector regression (SVR). A second result of this dissertation is a new optimization model that, ex ante, finds the best placement of a new wind farm, based on its overall positive impact for society. Utilizing historical data of wind power output for a set of established wind farms belonging to the Energy Reliability Council of Texas (ERCOT), this data-driven optimization problem explicitly considers the temporal and spatial dynamics of the wind power generation process via a predictive model built upon hybrid implementations of spatio-temporal random-effects models and random forests. The predictive model for wind power output is considered, per se, a significant contribution because it allows to conduct prospective analysis of the potential of a given location where no records are available and no farms have been installed. In total, three new models have been developed to better forecast and predict wind power production, in order to increase the effective integration of wind power in the electricity system.