Water Usage Pattern in Urban Areas

Abstract

Urban water demand forecasting is a critical component of short-term supply planning and real-time operational decision-making in smart city water management. This study presents a scenario-aware machine learning framework designed to predict daily urban water consumption across selected cities in Karnataka, India. The proposed framework integrates meteorological, temporal, and demand-related attributes to model consumption behavior effectively. A robust preprocessing pipeline is implemented, encompassing missing-value imputation, temporal sorting, lag and rolling feature construction, categorical encoding, and correlation-based feature selection. Three predictive models—XGBoost, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU)—are developed and compared. The optimal model is deployed via a Flask-based web application featuring SQLite storage, real-time weather data retrieval through Open-Meteo APIs, public holiday detection, prediction history management, and interactive scenario simulation capabilities for rainfall, heatwave, and holiday demand variations. The integrated framework offers water utilities a practical decision-support tool for proactive resource allocation and contingency planning. By enabling what-if analysis under extreme weather events and special calendar days, the system enhances municipal preparedness and contributes to sustainable urban water resource management. This work demonstrates the operational viability of combining machine learning with scenario-based planning to address the complexities of water supply systems in developing urban regions.