Aggregation strategies for grid-based Numerical Weather Prediction (NWP) to improve power curve models with meta-learning extension

Abstract

In the first part of this thesis, alternative strategies to estimate the wind power curve by utilizing grid-based weather forecasts from Numerical Weather Prediction (NWP) models are proposed. Traditional power curve estimation such as Weibull Cu mulative Distribution Function and Five Parameters Logistic Function uses actual wind speed information at a single location, and they are known to provide strong results. On the other hand, forecasts from multiple grid locations are used in a short-term wind power forecasting scenario which brings additional di. In order to resolve these, we propose a simple optimization framework which aggregates grid-based NWP predictions to estimate the power curve directly. Due to the problems with the cyclic nature of the direction, we propose alternative aggregation strategies based on generalized additive models. Our experiments on six wind farms show that parameter estimation with Quasi-binomial distribution assumption on the response provides superior performance compared to popular Gaussian likelihood assumption used in the power curve estimation approaches. In the second part, meta-learning approaches with a dynamic model ranking mechanism is applied. Three mostly used pointwise, pairwise and listwise ranking approaches are utilized with decision trees and tree-based ensembles as learners. Our experiments on the base-level model pool show that pairwise and listwise approaches improve the performance of both pointwise approach and the individual models.