As an important anthropogenic interference on the water cycle, reservoir operation behavior remains challenging to be properly represented in hydrologic models, thus limiting the capability of predicting streamflow under the interactions between hydrologic variability and operational preferences. Data-driven models provide a promising approach to represent reservoir operation rules by capturing relationships embedded in historical records. Similar to hydrologic processes vary across temporal scales, reservoir operations manifest themselves at different timescales, prioritizing different targets to mitigate streamflow variability at a given time scale. To capture interactions of reservoir operations across time scales, we proposed a hierarchical temporal scale framework to investigate the behaviors of over 300 major reservoirs across the Contiguous United States with a wide range of streamflow conditions. Machine learning models were constructed to simulate reservoir operation at daily, weekly, and monthly scales, where decisions at short-term scales interact with long-term decisions. We found that the hierarchical temporal scale configuration better captures reservoir releases than models constructed at a single time scale, especially for reservoirs with multiple operation targets. Model-based sensitivity analysis shows that for more than one third of the studied reservoirs, the release schemes, as a function of decision variables, vary at different time scales, suggesting that operators are commonly faced with complicated trade-offs to serve multiple purposes. The proposed hierarchical temporal scale approach is flexible to incorporate various data-driven models and decision variables to derive reservoir operation rule, providing a robust framework to understand the feedbacks between natural streamflow variability and human interferences across time scales.