Model Predictive Control for Battery Storage in Net-Load Levelling: A Comparison of MILP and Heuristic Strategies Under Perfect and Learned Forecasts

Abstract

The increasing integration of decentralised photovoltaic (PV) systems and electric vehicles into low-voltage (LV) distribution networks introduces significant variability in net-load patterns, exemplified by the “duck curve” (daytime overgeneration) and sharp peak demands. Addressing these challenges is critical to maintaining network reliability and minimising infrastructure costs. This study presents a comparative analysis of mixed integer linear programming (MILP) and heuristic control strategies within a model predictive control (MPC) framework for load levelling using a battery energy storage system (BESS), evaluated under perfect and LightGBM-based net-load forecast scenarios. These forecasts serve as input to the MPC framework, which optimally schedules MPC charging and discharging actions of MPC over a 24 h horizon to minimise the net-load deviations from predefined thresholds. Under perfect forecasts, MILP achieves superior load levelling performance through predictive optimisation, while the heuristic provides a robust, low-complexity baseline. With LightGBM forecasts, the performance gap between MILP and heuristic approaches widens due to forecast uncertainties, underscoring the critical role of robust control strategies in handling imperfect net-load predictions for BESS operation.