Incorporating risk attitudes in an irrigation reservoir management model

Thesis develops a technique whereby the attitudes towards risks of decision makers could be incorporated into a stochastic dynamic programming model of a reservoir by identifying the optimal release policy for the reservoir.. Development of optimal release policies for reservoirs with explicit recognition of the stochastic environment, e.g., uncertainty in future inflows, in which reservoirs must operate has been dominated by optimisation models which use summative measures of reservoir performance, e.g., expected benefits or costs. This study argues that these existing summative metrics do not adequately represent the desirability of a release strategy, and that they only constitute a nominally 'rational' decision making approach which may in fact be at odds with the true attitudes of the decision maker(s) towards the risk aspects of the decisions associated with that strategy. This situation is particularly relevant for reservoirs designed and operated to supply water to irrigation projects. In irrigation projects the possibility of individual farm bankruptcies arising from crop failure due to short-term water shortage must be balanced against maximising long-term benefits. A means for a more complete representation and evaluation of the possible consequences (likelihoods and magnitudes of a range of outcomes) associated with a release decision is therefore necessary. A model capable of replicating the manner in which risks associated with reservoir release decisions are perceived, interpreted and compared by a decision maker is proposed. The model is based upon Neural Network (NN) theory, and enables the more complete representation of the risk function of a particular decision to be considered in making decisions on reservoir releases. The neural network used to perform the decision making process is a feed-forward back-propagation network. The use of a more complete representation of the risks associated with the decisions in a release policy has been limited by existing optimisation frameworks. However this decision-making model is able to be used in a Stochastic Dynamic Programming (SDP) approach to give a Neural Network -Stochastic Dynamic Programming (NN -SDP) model. The resulting integrated model allows the attitudes towards risk of a decision maker to be considered explicitly in defining the optimal release policy. At each stage of the dynamic program, a risk curve is developed for each of the candidate release decisions. The neural network then selects the 'best' risk curve, and thereby the optimal release decision, through a pair-wise comparison process of these risk curves. Training of a neural network to perform this task is achieved using pairwise comparisons elicited from a decision maker a priori. The approach is demonstrated by application to the Fairbairn reservoir in Central Queensland, Australia. Release policies based on the risk preferences of three different decision makers were derived for this reservoir. Comparisons were then conducted against an optimal policy obtained using a SDP model with an objective function based on the more traditional expected squared deficit from target. The release policies reflecting the risk preferences of each of the three decision makers, and derived from the NN -SDP formulation, exhibited hedging and maintained the reservoir at levels higher than those associated with an operation policy defined solely by the expected squared deficit objective function. These results are consistent with previous studies on risk perceptions of decision makers which indicate that farmers tend to be risk averse. The NN -SDP model, and particularly its ability to consider a more comprehensive interpretation of risk in the development of water allocation and water release policies, is a significant step towards improving the ability of water resource managers to reflect and model risk more appropriately in the mathematical models used to identify optimal operating policies for reservoir systems.