Machine Learning Research And Application In Wind Forecasting

Manifold learning, also known as nonlinear embedding or nonlinear dimension-ality reduction, is an important research area of machine learning. The nonlinear em-bedding technique is to recover the intrinsic low-dimensional manifolds embedded in the high-dimensional space, which will be help to analysis and discovery the nature of data.In the past decade, a lot of nonlinear embedding techniques have been developed in a variety of literature. However, the quantitative evaluation criteria are less studied. The most commonly used criterion to estimate the embedding qualities of the nonlinear embedding algorithms in literature is the degree of neighboring preservation, which does not take the manifold related features into account. Although there are several other different evaluation ways, the quantitative criteria heavily depends on the prior knowledge to the functional form of the low-dimensional manifolds or geodesic dis-tances. In this sense, it is challenging but valuable to develop pervasive and specific criteria for the embedding qualities of the nonlinear embedding algorithms.In this thesis, we propose several novel criteria for quantitative evaluation, by con-sidering the global smoothness and co-directional consistence of the nonlinear embed-ding algorithms. Based on the concept of magnification factors (MF) and principal spread direction (PSD), the proposed criteria are geometrically intuitive, simple, and easy to implement with a low computational cost. Experiments show that our criteria capture some new geometrical properties of the nonlinear embedding algorithms, and can be used as a guidance to deal with the embedding of the out-of-samples.Except for theoretical research on machine learning, we also conducted practical applications in the new energy industry. The thesis concentrates on the integration of wind energy into the power grid, which requires accurate short-term prediction models for wind speed and direction.As a clean source of energy, renewable applications have gained considerable atten-tion over the past decade, for example driven by the ever growing international pres-sure of reducing the carbon footprint. This resulted in increased efforts by the academic research community as well as industrial R&D to examine the feasibility of wind and solar energy, as neither possess a direct carbon footprint and a limited indirect carbon contribution to the environment. What is more, as the events at Fukushima make for a powerful reminder of the vast dangers posed by nuclear power, we’ll see a global call to arms for the safest, most reliable energy sources:primarily solar and wind. Current-ly electricity is not directly storable in large amounts. And electricity production and consumption must be in balance at all times to reliably and safely operate an electrical grid. However electricity generated from wind power can be highly variable at sever-al different timescales. So technologies to manage the variability and intermittency of wind generation, including wind forecasting, is the key aspect associated to the optimal integration of that renewable energy into electricity grids.This thesis examines various model structures for identifying prediction models for wind speed and direction. The results of this analysis yield that causal autoregressive models that rely on meteorological data are to be preferred over noncausal ones, advo-cated in the recent literature. Furthermore, contrasting recent work using regression-based and classification formulations for predicting wind direction, the reported results show that the former technique produces a better performance. We also examine differ-ent modeling techniques, which yields that the random forest technique produces the most accurate prediction models. The reported comparisons are based on the analysis of recorded meteorological data from the Emirate of Abu Dhabi, U.A.E, between2002and2007.