Study On Reservoir Operation Based On Hydrological Forecast: Uncertainty Analysis And Optimization

Advances in weather forecasting, hydrologic modeling and hydro-climaticteleconnection relationships have improved hydrological forecasts considerably. Tocombine hydrological forecasts with optimization models provides a promisingapproach to improving reservoir system efficiency. Forecast uncertainty is a major issuein the use of hydrological forecasts in reservoir operation. On one hand, a short forecastmay not provide sufficient information; on the other hand, a long forecast can be toouncertain. As a result, effective forecast horizon is another important issue in reservoiroperation. The dissertation presents a systematic investigation of reservoir operationoptimization under hydrological uncertainty. Optimization models for reservoiroperation and statistical models for uncertainty analysis are set up, based on whicheffects of forecast uncertainty and hydrological uncertainty beyond forecast horizon onreservoir operation decisions are elaborated. The dissertation comprises three parts:Firstly, optimization algorithms are developed based on examinations of reservoiroptimization models and monotonicity property of operation decisions. With the focuson objective function, for water supply problems, the monotonic relationship betweenrelease decision and water availability is derived from the property of diminishingmarginal utility; for hydropower problems, a similar monotonic relationship is derivedthrough formulations of the complementary relationship between release decisions andreservoir storage. The monotonic relationship enables improvements of conventionaldynamic programming. An improved dynamic programming (IDP) algorithm and animproved stochastic dynamic programming (ISDP) algorithm are developed for watersupply problems. A successive improved dynamic programming (SIDP) algorithm isdeveloped for hydropower problems.Secondly, statistical models are set up to quantify and to simulate uncertaintyevolution in period-by-period updated hydrological forecasts. The model decomposestotal forecast uncertainty into forecast updates in intermediate periods and efficientlycaptures characteristics of forecast uncertainty, i.e.,“the longer the forecast horizon, thelarger the forecast uncertainty” and “forecast uncertainty reduces as time progresses”. Amartingale model of forecast evolution (MMFE) is set up to analyze unbiased andGaussian forecast uncertainties. In addition, biased and non-Gaussian characteristics of forecast uncertainties are illustrated based on real-world forecast data. To bridge the gap,a generalized marginal model of forecast evolution (GMMFE) integrating normalquantile transformation and MMFE is developed to simulate biased and non-Gaussianforecast uncertainties.Thirdly, operation decisions are analyzed considering the joint effects of forecastuncertainty and hydrological uncertainty beyond forecast horizon. While a shortforecast may not provide sufficient information and a long forecast can be too uncertain,the concept of effective forecast horizon (EFH) is to balance the confidence and amountof forecast information. The analysis shows that when there is a short forecast horizon,uncertainty of operation decisions is dominated by hydrological uncertainty beyondforecast horizon and to prolong forecast horizon leads to improvement of operationdecisions. When the forecast horizon is long, uncertainty of operation decision is insteaddominated by forecast uncertainty and to prolong forecast horizon even goes againstimproving operation decisions. The selection of effective forecast horizon is to balancethe marginal effects of forecast uncertainty and hydrological uncertainty beyondforecast horizon, aiming to minimize the uncertainty of operation decision.