A Comparison of the Performance of Advanced Statistical Techniques for the Refinement of Day-ahead and Longer NWP-based Wind Power Forecasts

Abstract:

Predictions from Numerical Weather Prediction (NWP) models are the foundation for wind power forecasts for day-ahead and longer forecast horizons. The NWP models directly produce three-dimensional wind forecasts on their respective computational grids. These can be interpolated to the location and time of interest. However, these direct predictions typically contain significant systematic errors (“biases”). This is due to a variety of factors including the limited space-time resolution of the NWP models and shortcomings in the model’s representation of physical processes. It has become common practice to attempt to improve the raw NWP forecasts by statistically adjusting them through a procedure that is widely known as Model Output Statistics (MOS). The challenge is to identify complex patterns of systematic errors and then use this knowledge to adjust the NWP predictions. The MOS-based improvements are the basis for much of the value added by commercial wind power forecast providers.

There are an enormous number of statistical approaches that can be used to generate the MOS adjustments to the raw NWP forecasts. In order to obtain insight into the potential value of some of the newer and more sophisticated statistical techniques often referred to as “machine learning methods” a MOS-method comparison experiment has been performed for wind power generation facilities in 6 wind resource areas of California.

The underlying NWP models that provided the raw forecasts were the two primary operational models of the US National Weather Service: the GFS and NAM models. The focus was on 1- and 2-day ahead forecasts of the hourly wind-based generation. The statistical methods evaluated included: (1) screening multiple linear regression, which served as a baseline method, (2) artificial neural networks, (3) a decision-tree approach called random forests, (4) gradient boosted regression based upon an decision-tree algorithm, (5) support vector regression and (6) analog ensemble, which is a case-matching scheme.

The presentation will provide (1) an overview of each method and the experimental design, (2) performance comparisons based on standard metrics such as bias, MAE and RMSE, (3) a summary of the performance characteristics of each approach and (4) a preview of further experiments to be conducted.