Probability Density Prediction of Wind Farm Power Generation: Benchmarking Natural Gradient Boosting Approach Using Ensemble Weather Forecast

Wind farm (WF) power generation fluctuates depending on the direction and speed of wind. Its prediction has been recognized as an essential elemental technology for grid connections of wind power generation. In particular, a quantitative understanding of the uncertainties in the prediction results is significantly important for designing the operational logic of energy management systems to compensate the unexpected behavior of WF power generation by means of alternative power sources and energy storage systems with appropriate capacities. This study focuses on the difficulties in the statistical learning of prediction models, which have hindered the realization of accurate probability density predictions (PDPs) using large-scale multidimensional data, such as the wind vector field derived using numerical weather prediction (NWP) models, in the PDP task of WF power generation, and discusses a prediction scheme that applies a nonparametric machine learning model based on the concept of natural gradient boosting, which is expected to contribute to overcoming this learning difficulty. This approach effectively uses high-dimensional information obtained from ensemble weather forecasting (EWF) and latest observations to achieve a plausible PDP of WF power generation. The agreement of the results of numerical experiments with data acquired from real-world sites demonstrate the effectiveness of the proposed framework.