This paper presents a grid-based model that aims to find a suitable spatial resolution to improve visualization and inference of the results of spatial load forecasting for feeders and/or distribution transformers. This approach can be considered as an unsupervised learning approach to cluster the input data (i.e., the power rating of the distribution transformers) in cells (clusters) to find a cell size that gives high internal homogeneity in the cells and high external heterogeneity of each cell with respect to its neighbors in order to reduce the inference errors that can affect the results of spatial load forecasting methods. The proposal was tested considering the spatial distribution of transformers installed in a real distribution system for a medium-sized city. Using the resolution determined by the grid-based model, it is possible to build a map of the spatial distribution of load density in a service area with a low relative local dispersion and a high relative global dispersion. To demonstrate the efficacy of the approach, spatial electric load forecasting of the study zone is performed using different spatial resolutions; the grid size determined via the proposed model represents the equilibrium between spatial error and computational effort, which is the main original contribution of this work. The techniques of spatial electric load forecasting are beyond the scope of this paper. © 2014 Elsevier B.V.