Development of methods for regional wind power forecasting

The large-scale integration of wind power can be a challenge for power system operators because, unlike conventional power sources, wind power is variable and non-dispatchable. To alleviate some of the problems posed by large-scale wind power integration, power system operators express the need for short-term (48 to 120 hours ahead) forecasts of the aggregated output of all wind farms within a specified geographical region. The aim of the thesis is to develop a framework and tools to help in the implementation of statistical regional wind power forecasting models. We first propose a framework for the characterization of the regional wind power. In this way, salient aspects of the regional wind power forecasting problem that must be taken into account when designing a regional forecasting model are identified. We then examine the regional forecasting problem from a statistical learning perspective. We define three generic approaches that can be used to combine sub-models to build regional models. The influence of these approaches on forecast accuracy is examined, as well as that of the choice of sub-models. The comparison of sub-models is made possible by the introduction of a novel forecasting model whose performance is shown to be comparable to that of other state-of-the-art models. Finally, we examine the impact of explanatory variable selection on forecast accuracy and derive general guidelines applicable in the frame of regional wind power forecasting. To ease modelling, automatic selection techniques are investigated. Two variable selection methods (a filter and a wrapper method) that exploit problem-specific characteristics are proposed. These methods are shown to compare very favourably to a generic state-of-the-art method.