Learning causality structures from electricity demand data

In this paper, we present an alternative approach to predictive modeling for future energy demands. It is based on the application of causal detection models to create specifications of how this demand might be caused by different environmental and social factors. We proceed by using a dataset generated by the wholesale electricity company of Argentina (CAMMESA) and selecting four prominent causal detection methods identified in the literature. These methods were selected based on their demonstrated effectiveness and widespread adoption. Since these causal detection methods yield different causal graphs, we were able to construct an ensemble model that achieved better performance for recovering the true causal structure when applied to the full dataset. Also, we show that the variables in the causal model can be used to yield more accurate forecasts of future demands, improving over the informal models used by staff in electricity utilities.