Study region: The Xiaolangdi Reservoir in the Yellow River Basin, China. Study focus: Static drought-limited water levels (DLWLs) often fail to address the non-stationary nature of prolonged megadroughts. To address this gap, we derived dynamic, stage-dependent DLWLs that evolve alongside the drought lifecycle. Instead of treating drought as a uniform event, our method splits the process into four distinct evolutionary phases-Gradual Emergence, Escalation, Persistence, and Recovery-based on observable shifts in meteorological signals. We employed a supervised Random Forest (RF) model, trained on a physically constrained Comprehensive Drought Index (CDI), to grade severity. Crucially, we contrasted this approach against an unsupervised K-means clustering baseline to test whether statistical patterns alone could capture the necessary physical drought signals. New hydrological insights: Our analysis shows that the supervised model far exceeds the unsupervised baseline in reliability (AUC=0.98; RMSE=0.62). This result is significant because it proves that operational tools must be anchored in physical benchmarks-specifically those capturing soil moisture memory-rather than relying on latent statistical clusters. The primary contribution here is a structured set of dynamic DLWL t curves. We found that the operational implication of a specific drought severity is contingent upon the timing: a "severe" signal demands preemptive hedging during the emergence phase but survival-based conservation during persistence. By aligning reservoir decisions with these specific evolutionary stages, this approach offers a practical pathway to improve resilience in semi-arid basins.