Deriving representative reservoir operation rules using a hidden Markov-decision tree model

Reservoirs have been widely used to regulate streamflow to meet both human and natural water requirements. This study applies a hidden Markov-decision tree (HM-DT) model to derive representative reservoir operation modules under various operation conditions (i.e., inflow, storage, as well as unknown factors) and their transitions (the dynamic change of operation rules) that reflect the impacts of seasonality, long-term non-stationarity, and extreme events on reservoir operation. The representative operation modules can be applied to reservoirs in the same region that are not observed; the capability for simulating dynamic operation behaviors improves the predictive accuracy as compared to regular decision trees. Using a number of reservoirs located in the same region for training, the HM-DT model can derive a limited number of representative operation modules in the form of decision trees (DT), and the transitions between different operation schemes in response to changing operation conditions. The application of the HM-DT model is demonstrated through a case study of the Upper Colorado River basin, where eight representative operation modules are determined for 50 reservoirs located in the region, and the modules are validated with 11 reservoirs in the same region. The eight operation modules are classified into three types (i.e. nearly constant release, release as a piece-wise function of inflow, and release almost identical to inflow). The identified operation modules and the transition patterns between operation modules can be used to better understand real-world operation behaviors, improve future operations, and build realistic reservoir operation components in basin-scale hydrological models.