Investigating LSTM for DMA Hourly Demand Prediction: A novel Decomposition-Enhanced Approach

Accurate short-term demand forecasting in District Metered Areas (DMAs) is essential for enhancing pressure control and minimizing non-revenue water losses. Long Short-Term Memory (LSTM) networks are widely used for this purpose due to their ability to capture nonlinear temporal dependencies. However, existing research faces two major limitations: (1) insufficient systematic evaluation of LSTM parameter sensitivity across diverse DMAs, leading to performance instability, and (2) the excessive complexity of hybrid decomposition methods, which impedes their practical implementation in real-world settings. To address these issues, this study proposes a physics-informed Decomposition LSTM (D-LSTM) framework that employs synchronous averaging to extract periodic patterns, allowing the LSTM to focus solely on irregular components—eliminating the need for complex signal processing while enhancing predictive accuracy. The framework is validated on five DMAs with varying characteristics under 56 parameter combinations. Results indicates that: (i) demand characteristics have a greater impact on accuracy than parameter tuning; (ii) D-LSTM consistently outperforms standard LSTM, reducing MAPE by up to 8.6% in low-demand DMAs; and (iii) the model remains robust even with suboptimal parameters and avoids specialized decomposition methods, providing a practical and accessible solution for water utilities with limited technical resources.