Short-term Wind Power Forecasting Approach Based On Gaussian Process Regression And Trust-Tech

The stochastic nature of wind power has posed a significant challenge to the operation of a power grid.Key issues such as the system operation,electricity market price,and requirements of spinning reserves should be reconsidered to incorporate the impact of wind power.All of these applications depend on the outputs of short-term wind power forecasting,so high-accuracy prediction is essential.Supervised learning models using numerical weather prediction(NWP)data have been utilized for short-term wind power forecasting tasks and only involve NWP data of the target wind farm as essential features.This dissertation proposes a novel method to improve conventional supervised forecasting models.It extends supervised forecasting models to semi-supervised forecasting models.The method incorporates extended numerical weather(ENW)data as feature inputs and then takes advantage of unlabeled numerical weather data to utilize these features properly.The key step of the proposed method is a data-driven approach.It builds a proper feature extraction for the ENW features and provides supplementary features.The only modification to the existing supervised forecasting model is the addition of these supplementary features,which are easily implemented.For illustrative purposes,the Gaussian process(GP)is used as the supervised forecasting module to be improved.The training of a GP is a non-linear optimization problem.Each local optimal solution corresponds to an interpretation of the data.In this dissertation,the Trust-Tech(TRansformation Under STability-reTaining Equilibria CHaracterization)global optimization approach is applied to compute multiple optimal solutions.Given multiple interpretations,a model average is used to improve the performance of the model.The computation burden of multiple solutions is heavy.In order to reduce the computational burden,a bi-level approximation method is proposed.The time complexities of the related methods are analyzed.Compared to the direct application of tier-1 Trust-Tech approach in the training of a GP,the proposed method requires less computation.In this dissertation,a high-accuracy point forecasting model,based on a Gaussian process that is a kernel machine,is proposed.One distinguished feature of the proposed model is that prior knowledge of NWP is encoded into the kernel building for high forecasting accuracy.In addition,the issue of multiple local optimal solutions during the training process is explored by the Trust-Tech optimization method to further improve the forecasting performance.The performances of the proposed methods are evaluated on the 2012 global energy forecasting competition(GEF Com 2012)wind power forecasting data which are public available.The proposed forecasting model outperforms the best published result.