Li
Battery energy storage is central to resilient, renewable microgrids, and increasingly to the reuse of retired electric vehicle batteries and the grid integration of large new loads such as AI data centers. Long-term planning models must simplify how batteries charge, age, and operate to remain solvable over decades, yet each simplification can change the size, cost, or reliability of the resulting design. This dissertation develops a series of optimization frameworks that test whether such simplifications preserve the planning conclusions they support.
The research first addresses how planning models represent battery operation. An iterative correction framework aligns the battery degradation cost with the depth of discharge the optimized schedule actually produces, improving cost accuracy while keeping the model linear and at full temporal resolution. State-of-charge-dependent charging limits then show that realistic, tapered charging calls for larger storage, and that designs sized under constant-power charging can fail to serve load once tapering occurs. The planning horizon is then extended to the battery’s full life. A plan–validate–refine framework reveals energy shortfalls that emerge only late in service and shows that second-life batteries must be compared with new ones at equal reliability, where they remain competitive only at a substantial price discount. A controlled modeling fidelity study further shows that models agreeing closely on lifecycle cost can still misrepresent energy adequacy, and that representative day methods can miss the critical shortfall event.
The final studies move to the pack and grid-interface levels. A robust cell selection and assignment model assembles compliant packs from heterogeneous retired cells under measurement uncertainty, where single metric sorting fails. A developer-side planning framework for AI data centers treats grid import capacity as a decision, showing that onsite photovoltaics, batteries, and flexible workloads can meaningfully reduce the capacity a site must request, and that workload flexibility mainly substitutes for battery investment. Together, these contributions show that the appropriate level of model detail depends on the decision being made: operating assumptions change storage size, agreement in cost does not imply agreement in reliability, and joint, robust decisions outperform single-metric rules.