Model selection criteria for short-term microgrid-scale electricity load forecasts

The electricity grid is evolving from a monolithic centralized system to a smart distributed system, composed of distributed and renewable generation resources, where power supply and demand balancing is needed at a microgrid scale. In this paper, we explore model selection criteria for short-term microgrid-level load predictions. To this end, we experiment with five different models in the context of usage traces from six diverse sites collected over a period of eight months. We find that model selection is heavily influenced by the variability in the data and that models which do not use weather forecast information but rely only on historical usage data perform better on sites with highly variable loads.