Learning Distance Effect on Lignite Quality Variables at Global and Local Scales

Determining scale and variable effects have critical importance in developing an energy resource policy. This study aims to explore the relationships in heterogeneous lignite sites using different scale models, spatial weighting as well as error-based pair-wise identification. From a statistical learning framework, the relationships among the quality variables such as geochemical variables and the contributions of the coordinates to quality measures have been exhibited by generalized additive models. In this way, the critical roles of spatial weights provided by the coordinates have been specified at a global scale. The experimental studies reveal that incorporating the geological weighting in the models as the additional information improves both accuracy and transparency. Because relationships among lignite quality variables and sampling locations are spatially non-stationary, the local structure and interdependencies among the variables were analyzed by geographically weighting regression. The local analyses including spatial patterns of bandwidths, search domains as well as residual-based areal dependencies provided not only the critical zones but also availability of pair-wise model alternatives by calibrating a model at each point for location-specific parameter learning. The results completely show that the weighting models applied at different scales can take spatial heterogeneity into consideration and these abilities provide some meta-data and specific information using in sustainable energy planning.