UC San Diego

Grid-connected battery energy storage systems (BESS) are a promising technology to decrease peak load. In this paper, a linear programming (LP) optimization model of a BESS that minimizes demand charge in real-time is developed. The model is employed to study the effect of two temporal resolutions of load profiles, 15~min and 1~hour, on peak load shaving and optimal sizing of BESS. In contrast to previous studies with perfect forecast, realistic forecast errors are prescribed; a 24~hour persistence forecasts and a recurrent neural network (RNN) model are used for short-term load forecasts. The LP routine was coupled with model predictive control (MPC) to obtain optimal BESS dispatch schedules and to reduce the deviations due to load forecast errors in real time. The performance of the proposed algorithm is simulated for one month on a grid-connected building. Compared to perfect load forecasts, the RNN forecast error causes a monthly average of 1.3~kW higher optimized net load, and corresponding \$27 higher optimal demand charge (ODC). Correcting persistence forecast with more accurate RNN forecast yield better peak shaving performance with a monthly average of 28.1% reduction in optimized net load. The temporal load resolution has a significant influence on the ODC. The difference of ODC for two temporal resolutions depend strongly on the load profile characteristics. When the battery capacity is above a certain power and energy threshold limit, the 1~hour load profile is found to be sufficient for determining ODC. However, when the power and/or energy is constrained, the difference in the ODC between 15 min and 1 hour resolution ranges from $14.7 to $123.1 for 28 days.