Modeling and Comparing GMDH and RBF Artificial Neural Networks in Forecasting Short-Term Demand for Drinking Water in Zahedan

The water requirement of the urban district of Zahedan City in Iran is met by the transfer of water from Sistan Chah-Nimeh reservoirs that are themselves suffering from severe water crisis. So, the prediction of the urban demand for drinking water can greatly help managers and users of urban water systems to use sound management practices. The present paper uses the artificial neural networks of GMDH and RBF, as two vigorous tools for the analysis and modeling of nonlinear relationship, to estimate monthly demand for drinking water in Zahedan in 2017. The selected parameters include mean monthly temperature, relative humidity percentage, mean precipitation, sunny hours, and previous-month consumption rate. The results of comparing MSE and MAE indicators show that after studying different structures with various number of neurons and hidden layers, the GMDH neural network with three hidden layers that has one neuron in layer 1, three neurons in hidden layer 2, and three neurons in hidden layer 3 yields the best results for the predication of short-term demand for drinking water. The comparison of linear and nonlinear activity functions reveals that in output layer of GMDH and RBF neural models, the nonlinear functions outperform linear functions. As well, among the GMDH model, the models with nonlinear output show better performance than those with linear output. Also, it is shown that the expansion of the network structure cannot improve the results considerably.