A CEEMDAN-Transformer-BiLSTM Framework for Multi-Scale Urban Water Demand Forecasting

Accurate forecasting of urban water demand is essential for scientific regulation and sustainable management of water resources, particularly in complex DMA (District Metered Area) environments. This study proposes an integrated regional water demand prediction framework that combines CEEMDAN decomposition with deep learning techniques. CEEMDAN is first applied to decompose the original water demand time series into multiple Intrinsic Mode Functions (IMFs), effectively extracting multi-scale features and mitigating non-stationarity and complexity. A hybrid Transformer-BiLSTM model is then constructed to capture global dependencies, nonlinear dynamics, and bidirectional temporal features. Experimental results demonstrate that the proposed CEEMDAN-Transformer-BiLSTM model significantly outperforms various benchmark models in terms of prediction accuracy, robustness, and generalization across different DMAs. This research provides a new perspective for modeling complex water resource time series and offers theoretical and practical support for optimizing urban water allocation and achieving sustainable management, while laying a foundation for future work involving external driving factors, enhanced model interpretability, and dynamic regulation mechanisms.