Probabilistic forecast models for hydro-environmental characteristics and risk-based adaptive reservoir operation

Abstract

This study was motivated by the desire to improve risk-based decision making and adaptive management of large-scale water resources systems centering on multi-reservoir system operation. Forecasting the dynamic behavior of a water resources system is inherently uncertain. Case studies were performed using data from the Geum River basin in Korea. The overall objective of the research was to develop a methodology for managing a water resources system in an adaptive manner accounting for risks and uncertainties of the hydro-environmental characteristics. The characteristics considered in this research are the stage-discharge relationship, reservoir inflow, and water qualities in terms of biological oxygen demand (BOD) and total phosphorus (TP). First, stage-discharge ratings were developed and assessed using both deterministic and probabilistic methods at two stage-discharge measurement stations chosen because they exhibited hysteresis. For deterministic approaches, nonlinear programming (NLP), fuzzy rule-based modeling, and a one-dimensional hydrodynamic model were used. For the probabilistic approach, a Bayesian Markov chain Monte Carlo (MCMC) technique was employed. Based upon a comparison of the methods, a hybrid methodology which combines NLP and Bayesian MCMC was proposed as the best alternative. Second, stochastic monthly inflow forecast systems were developed using stochastic artificial neural networks and nonparametric modeling. To determine whether or not a k-nearest neighbor (k-NN) bootstrap resampling method might be used in practice for daily inflow forecasts aimed at short term reservoir system operation, a daily forecast model was developed. In the context of practical applicability, it was concluded that the k-NN method was preferred due to its ease of application. In addition, it was demonstrated that this method can be applied successfully for daily inflow forecasting. Third, probabilistic BOD and TP models were developed using Bayesian networks. The relationships between reservoir release and risk of violating the water quality standards were derived. The case study clearly demonstrated that the probabilistic models overcome the weaknesses of deterministic water quality models by offering information about risks of violation of standards or failures to meet targets. Compared to the other methods for uncertainty analysis such as sensitivity analysis, first order second moment (FOSM) analysis, and Monte Carlo methods, the advantages of a Bayesian MCMC technique were identified. Fourth, instead of relying on the classical rule curves for reservoir system operation, an adaptive sampling implicit optimization (ASISO) model was developed that considered multiple objectives of energy production, water supply, and water quality management in terms of BOD and TP. A decision support system (DSS) especially designed for interactively integrating the probabilistic inflow, BOD and TP forecast models to the ASISO model was developed. The ASISO based DSS demonstrated an alternative for reservoir system operation by combining simulation and optimization algorithms and incorporating the risk of water quality standard violation and adaptive sampling of the inflow series. The case study also showed that the reservoir inflow forecast systems played a very important role in terms of differences between the models considered. This research contributed to implementation of adaptive reservoir system operation with consideration of risk and uncertainty by joining probabilistic forecast models for hydro-environmental characteristics to reservoir system operation, which has been considered a very daunting task. Probabilistic forecast models were proposed by comparing several popular methodologies. This research showed the possibility of application of stochastic artificial neural networks (ANNs) in the field of water resources. It is recommended for further study that the developed reservoir operation system be reduced to a weekly or daily time step. For sophisticated short term inflow forecast models in addition to the k-NN method, Markovian autoregressive models should be investigated by incorporating a variety of exogenous variables such as temperature and humidity.