Machine Learning and Water Economy: a New Approach to Predicting Dams Water Sales Revenue

The proper prediction of water sales revenue allows for pricing policies with a specified trend for the optimized use of water resources. The present work focuses on the prediction of the economic status of water sales revenue in a semi-arid environment. To meet this objective, evaporation data (E), dam input water volume (I), and dam output water volume (O) are used as independent factors to estimate water revenue (R) in the case study of Jiroft Dam, Iran. Different machine learning models are used, including classification and regression tree (CART), Chi-squared automatic interaction detector (CHAID), multi-layer perceptron neural network (MLP), and radial basis function neural network (RBF). The data are obtained daily from 20 March 2012 to 20 March 2015 and defined in six input combinations to the models using multicollinearity analyses. To compare these models, the Nash-Sutcliffe efficiency coefficient (NSEC), the root mean square error (RMSE), and the coefficient of correlation (CC) criteria are employed. All the models act better when records of water sales revenue are incorporated as additional input factors to the machine learning models. The MLP neural-based model indicates the best predicted values for daily water sales revenue (RMSE = 638.3 $ and CC = 0.798) followed by the RBF neural model (RMSE = 655.1 $ and CC = 0.786).