An Ensemble-Based Approach for Water Demand Prediction: Comparative Analysis of Conventional, Machine Learning, and Deep Learning Techniques

Water demand prediction is vital for effective infrastructure planning and sustainable water resource management. Accurate forecasting enables better decision-making in resource allocation and conservation efforts, ensuring equitable access to water. While traditional and machine learning approaches have shown promise in this domain, there is a growing need for methods that combine the strengths of these techniques to achieve higher predictive accuracy and adaptability. In this study, we propose a novel Ensemble Learning method that integrates Seasonal Autoregressive Integrated Moving-Average with Exogenous Regressors (SARIMAX), Support Vector Regression (SVR), and Artificial Neural Networks (ANN) using a stacking approach. This method is specifically designed to leverage the complementary strengths of statistical, machine learning, and deep learning models for improved water demand prediction. We compare the proposed method against 14 other predictive models, including classical approaches like Autoregressive (AR), ARIMA, and SARIMAX, as well as advanced machine learning and deep learning models such as Random Forest (RF), Extreme Gradient Boosting (XGBoost), LightGBM, Categorical Boosting (CatBoost), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM). The models are evaluated on a real-world dataset of monthly water consumption from Istanbul municipality spanning 2011–2023. Performance is assessed using multiple metrics, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). Results demonstrate that our proposed Ensemble Learning method consistently outperforms the other models, showcasing its superior predictive accuracy and robustness for water demand forecasting.