A comparison between neural network approach and multiple regression procedures for deriving reservoir operation policies

An optimisation model, based on discrete differential dynamic programming was applied to compute optimal monthly releases from a reservoir by adopting the sum of the squared deficit as objective function. A reservoir supplying an irrigation district in Sicily was used as a case-study. Operating rules were derived by interpolating the results of the DP model on a 32 years training period using both multiple linear regression and neural network techniques. Reservoir storage at the beginning of the current month, streamflow and release during the previous month were selected as independent variables and monthly linear multiple regressions were used to compute release for each month. Also, neural networks with 1 hidden layer, 1 neuron in the output layer, and different neurons in the input and hidden layers were applied by using a logistic activation function, and a network with 3 neurons in the hidden layer was preliminary selected on the basis of the minimum root-mean-square error. However, the comparison among different neural networks based on the simulated performance of the reservoir, led to prefer a simpler network with an identity activation function during July-September period. Then the best operating rules determined by the two techniques (multiple regression and neural network) were compared by simulating the reservoir performance during a 4-year validation period. The comparison was also extended to include the results of the Dynamic Programming in such period (under the hypothesis of perfect knowledge of future streamflows) and the results of the reservoir simulation by using the Standard Operating Policy (SOP). Results show that the use of operating rules based on the neural network approach leads to better performance of the reservoir both in terms of sum of squared deficits and of maximum monthly deficit.