Derivation of Operation Rules for an Irrigation Water Supply System by Multiple Linear Regression and Neural Networks

An approach based on optimization-multiple regressions or neural networks is applied for deriving the operating rules of an irrigation supply reservoir. Two Dynamic Programming models are adopted to determine the optimal releases by using objective functions based on the sum of squared deficits, subject to various constraints. The two models differ for the presence of a penalty term enforcing solutions which consider system management criteria, based on social and political constraints. Then, monthly operation rules are derived by expressing the optimal releases as functions of significant variables by regression or neural networks approach. Both regression equations and neural networks parameters are derived from optimization results on a long period including severe drought events; this period is also used to select the best rules. Then, the behaviour of such operation rules is assessed on a different shorter period, by simulating reservoir operation and computing several performance indices of the reservoir and crop yield through a soil water balance model. Results show that neural networks approach appears to improve the reservoir operation and that operating rules based on optimisation with constraints resembling real system management criteria, yield good performance both in normal and in drought periods, reducing maximum deficits and water spills.