베이지안 네트워크와 딥러닝을 이용한 안동댐의 월 유입량 예측 및 위험 회피 저수지 운영 방안

Abstract

Pragmatic reservoir operation providing for droughts needs an appropriate operating rule and the prediction of future inflow. For this reason, a method of hedging policy on reservoir operation proposed in this research includes a derivation of discrete hedging rule curves applying the mixed-integer programming and a monthly inflow prediction method using a copula-based Bayesian network combined with deep learning. The hedging stages consist of concern, caution, alert, and severe that accompany the required flow release, instream flow reduction, agricultural water rationing, and municipal water diminishing, respectively. Each derived hedging rule curve was decided by twelve monthly trigger volumes of available water for rationing for the respect drought stage. The available water for a reservoir is the current storage plus the inflow volume of a future operation period. The reservoir operation applying the discrete hedging rule requires a reliable prediction inflow. The two methods yielded the prediction inflow: one was a copula-based Bayesian network (CBN) to forecast the probability density function (PDF) of the monthly inflow; the other one was deep learning to forecast effective drought index (EDI). The procedures to determine the predicted inflow are as follows. One converts the PDF of the predicted inflow to the standard normal distribution function. Assume that the predicted EDI value is equal to the random variate of converted monthly inflow. Calculate the non-exceedance probability of the transformed random variate of the monthly inflow. Finally, one finds the monthly inflow with the same non-exceedance probability on the original PDF. Reservoir simulations utilized the forecasted inflow. The study applied the methods mentioned above to Andong dam reservoir. Using the reservoir inflow records from 1980 to 2013, The study derived the hedging rule curves of reservoir operation, constructed the CBN model, and trained the deep learning model. For the verification period from 2014 to 2018, the combined CBN with deep learning model forecasted the reservoir inflow, and the reservoir simulation was followed using the forecasted inflow. Comparing the forecasted inflows between the mode of the PDF of the predicted inflow using CBN and the combined CBN and deep learning model, the two prediction methods' forecasted results showed similar performance for the whole validation period. For the drought period of 2015, however, the forecasted inflow results from the combined CBN and deep learning model improved Nash-Sutcliffe efficiency by 6.43 and percent bias by 48% compared with the mode of predicted inflow PDF using CBN. The reservoir simulation results applying the hedging rule curves and the predicted inflow were 156% better in terms of temporal reliability and 19% better in volumetric reliability than the historical records. The more reliable results from the reservoir simulation were outcomes of the role of the forecasted inflow: even if the current storage is small, the events of large forecasted inflows made it possible for the storage phase to recover to the normal phase rapidly.