WaterPrompt-LLM: A single global model with a prompt pool for transferable cross-DMA water demand forecasting

Short-term water demand forecasting (WDF) across heterogeneous district metered areas (DMAs) supports fine-grained operation of water distribution systems. Developing a single global model remains challenging because DMAs differ in demand scale, user composition, and temporal patterns. DMA-specific models cannot share information, whereas conventional global models struggle with contrasting demand patterns. This study proposes WaterPrompt-LLM, which performs feature-guided prompt selection from a pool shared across DMAs. Demand features are used to select a soft prompt for each forecast. The selected soft prompt is concatenated with the temporal input and passed to a frozen pretrained large language model (LLM) backbone. On the Battle of Water Demand Forecasting (BWDF) benchmark, covering 10 DMAs and 40 DMA-week tasks, WaterPrompt-LLM achieved favorable results among the BWDF top-10 submissions. Task-level robustness was its main strength. It was one of only two methods without a negative-R² task. Its mean minimum task R² was 0.179, compared with 0.046 for Rank 2, and its mean absolute error (MAE) interquartile range was the lowest at 0.580 L/s. Across ten zero-shot tests, each excluding one target DMA from model development, the symmetric mean absolute percentage error (SMAPE), averaged across the ten tests, was 6.484 % and the mean R² was 0.729, close to the supervised BWDF Rank 3 values of 6.545 % and 0.739. All eight prompts were selected in multiple DMAs, with assignments associated with demand magnitude, within-week variation, and periodic shape. These results show that WaterPrompt-LLM provides a practical route to robust cross-DMA forecasting and transfer to unseen DMAs.