Evaluating the performance of feature selection techniques and machine learning algorithms on future residential water demand

The provision of potable water as a severe challenge has engaged many people worldwide. So, identifying influential factors in water demand forecasting (WDF) for the residential sector performs a vital role in water crisis management. Nowadays, long-term macro planning for vast geographic areas helps policymakers to achieve sustainable development goals. This study uses the same perspective to present a pattern for water consumption behavior and prediction. For this purpose, yearly residential water consumption data, along with climatic characteristics, and socioeconomic factors of rural areas of Isfahan, Iran are aggregated. The feature selection task is conducted on the collected data using various machine learning (ML) methods along with a novel approach, forward selection based on smoothness index (FSSmI). Posterior to selecting features influencing residential water demand (WD), the raw data are analyzed using regression techniques, including multiple linear regression, support vector regression, and random forest regression. The employed methods show an improvement in the feature selection procedure and coefficient of determination as a result of implementing the FSSmI method. Based on the results, the multiple linear regression and support vector regression gain 96 and 95% accuracy and less than 11 and 13% error respectively; it demonstrates the validity of forecasting methods.