Rapid urbanization and shifting demographic patterns are placing unprecedented strain on urban water and energy systems, particularly in densely populated and climate-sensitive regions. Effective urban planning requires predictive insights into how future population growth will influence resource demand. This study proposes a novel machine learning (ML) framework for forecasting water and energy requirements driven by population dynamics, with the goal of enabling smarter, data-informed urban growth strategies. The framework integrates historical demographic trends, spatial distribution, socioeconomic indicators, and climate variables to generate high-resolution forecasts of population change. These forecasts are then linked to predictive models for water and energy consumption using a hybrid ML architecture that combines Long Short-Term Memory (LSTM) networks with Extreme Gradient Boosting (XGBoost) to capture both temporal dependencies and nonlinear relationships. Model training and validation were conducted using data from a rapidly urbanizing metropolitan region over a 15-year span. Results demonstrate significant improvements in forecast accuracy, with R² values exceeding 0.92 for water demand and 0.89 for energy use, outperforming conventional statistical and standaloneMLmodels. The framework also enables scenario-based simulations, helping planners evaluate the impacts of different growth, climate, or policy conditions on resource infrastructure. Beyond its technical performance, this research provides a replicable model for integrating AI-driven forecasting into urban decision-making processes. It empowers policymakers and city planners to anticipate service bottlenecks, optimize infrastructure investments, and enhance urban resilience against climate and demographic shocks. The study concludes with recommendations for embedding such intelligence frameworks within smart city ecosystems, reinforcing the value of population-centric planning in achieving long-term urban sustainability and resource equity. We frame a population-informed residual hybrid that couples LSTM long-range dynamics with an XGBoost error-corrector trained on out-of-fold residuals, evaluated under rolling-origin splits with compute-aware deployment notes (CPU-only/quantized) to enable practical adoption.
Jubilson et al. (Thu,) studied this question.