Stochastic weather inputs for improved urban water demand forecasting: application of nonlinear input variable selection and machine learning methods