Examination of Spatial Relationships Using Machine Learning Techniques

The analog ensemble (AnEn) is a computationally efficient approach to generate accurate predictions and appreciable uncertainty quantification for a range of applications. Literature has shown the AnEn method produces significant improvements in forecasting for air quality, wind power production, and solar energy production, among other fields.  Further improvement is reached by optimally weighting the predictor variables used in the similarity metric.  However, in the literature this optimized weighting has only been performed for a select number of locations with a specific application (i.e., wind power generation) and spatial relationships have not been investigated.

 This research addresses two primary objectives using machine learning techniques to understand relationships present within the optimal predictor weighting determined and within the ensemble of analogs generated using the AnEn method. First, to test and validate an application of the AnEn method with optimized predictor weighting across numerous forecast locations. Specifically, short-term prediction of meteorological variables (i.e., wind speed, temperature) are generated at 669 METAR stations across the continental U.S. A brute force algorithm is used to test multiple weighting combinations of the predictor variables.  Results show that certain variables exert greater influence in determining the similarity, and specific weighting schemes lead to more accurate analogs. Second, machine learning classification is used to characterize the spatial relationships present within (a) the optimal weighting strategy and (b) the ensemble of analogs generated at each site.  The long-term objective is to use these results to guide the extension of the AnEn method to gridded fields. The primary benefit of the proposed research is the ability to quantify and justify uncertainty in a computationally efficient manner.