Optimal dispatch of battery energy storage on solar-dominant distribution grids requires accurate time-series characterization of net loads at all hours of the day and over multiple seasons to mitigate both peak loading and to help manage reverse power flow during periods of high solar production / low demand, such as sunny spring and fall days.
We compared methodologies for predicting time-series loads in commercial and industrial (C&I) buildings and present results from a 16-month pilot in which the best-performing prediction methodology was used, together with energy storage and solar PV, to shape load on a renewables-dominant distribution grid. Initial evaluation used facility-specific customer-provided process schedules in combination with historic load data to generate predictions based on site-specific heuristics. Field trials highlighted several shortcomings, including (1) inaccuracies in customer-provided schedules; (2) lack of timely access to schedule information; and (3) level of effort required to develop detailed process knowledge. Subsequent parametric testing was used to evaluate the sensitivity of prediction accuracy while varying prediction methodology and predictors, with emphasis on identifying predictors that can be auto-calibrated with minimal site-specific knowledge.
Our analysis showed that a regression-based approach using site-agnostic predictors offered prediction accuracy of 5% root-mean square error (RMSE) for 1-hour ahead predictions to 10% RMSE for 24-hour ahead predictions over the course of the 16-month trial. Notably, including site-specific production schedules (approximately 5-8% RMSE) offered only marginal improvements in accuracy relative to the site-agnostic approach, while requiring significantly more effort to implement and maintain.
Read the full paper here: Link