Building more realistic reservoir optimization models using data mining – A case study of Shelbyville Reservoir

In this paper, we promote a novel approach to develop reservoir operation routines by learning from historical hydrologic information and reservoir operations. The proposed framework involves a knowledge discovery step to learn the real drivers of reservoir decision making and to subsequently build a more realistic (enhanced) model formulation using stochastic dynamic programming (SDP). The enhanced SDP model is compared to two classic SDP formulations using Lake Shelbyville, a reservoir on the Kaskaskia River in Illinois, as a case study. From a data mining procedure with monthly data, the past month’s inflow (Qt−1), current month’s inflow (Qt), past month’s release (Rt−1), and past month’s Palmer drought severity index (PDSIt−1) are identified as important state variables in the enhanced SDP model for Shelbyville Reservoir. When compared to a weekly enhanced SDP model of the same case study, a different set of state variables and constraints are extracted. Thus different time scales for the model require different information. We demonstrate that adding additional state variables improves the solution by shifting the Pareto front as expected while using new constraints and the correct objective function can significantly reduce the difference between derived policies and historical practices. The study indicates that the monthly enhanced SDP model resembles historical records more closely and yet provides lower expected average annual costs than either of the two classic formulations (25.4% and 4.5% reductions, respectively). The weekly enhanced SDP model is compared to the monthly enhanced SDP, and it shows that acquiring the correct temporal scale is crucial to model reservoir operation for particular objectives.