Cross-Sector XGBoost Model Sharpens Regional Water-Demand Forecasts in Taizhou, China

Reliable forecasts of how much water households, farms, factories and ecosystems will need are central to regional water-network planning, drought preparedness and demand-side scenario analysis. Yet conventional approaches typically model domestic, agricultural, industrial and ecological demands separately, leaving the historical associations among these categories underused. A new study addresses that gap with a Cross-Sector XGBoost framework for regional multi-sector water-demand forecasting, built from a water–food–energy–ecology nexus perspective.

The framework represents regional demand through six management-relevant categories: domestic and public use, grain irrigation, other agricultural uses, energy-intensive and strategic industrial use, general industrial use, and ecological and environmental demand. It augments sector-specific predictors with candidate cross-sector historical features, and uses training-period expanding-window validation to select which cross-sector representation to retain, all while maintaining a region–sector–month forecasting structure. The method was tested with monthly water-demand records from Taizhou, China, and benchmarked against seasonal-average, Holt–Winters, LSTM and Independent XGBoost models.

The results showed measurable but nuanced gains. The selected Cross-Sector XGBoost reduced root-mean-square error from 95.57 to 87.89 × 10⁴ cubic metres — an 8.04% improvement — and lowered weighted mean absolute percentage error from 12.62% to 11.82% relative to Independent XGBoost. Lower RMSE was achieved in five of the six demand categories, and cluster-bootstrap analyses supported the improvement on the fixed data split. Feature-group ablation revealed that the predictive value of cross-sector information depended strongly on how it was represented, underscoring that simply adding more inter-sector variables does not guarantee better forecasts.

The study also reports an important caveat. In annual walk-forward evaluation, the framework showed no aggregate RMSE improvement under sequential updating, indicating temporal variability in the benefit of cross-sector information. In other words, the advantage of borrowing historical signals across demand categories is not constant over time, and the authors position the framework as a way of selectively incorporating such information rather than as a universal upgrade. By structuring which cross-sector features are retained, the approach provides demand-side information for regional water-resource management while remaining transparent about when that information helps and when it does not.

Subject of Research: A machine-learning framework that uses cross-sector historical information to forecast regional multi-sector water demand.

Article Title: Regional Multi-Sector Water-Demand Forecasting with Cross-Sector Historical Information

Article References: Zhang, Y., Dong, L., Li, X., Fang, G., & Zhong, H. (2026). Regional Multi-Sector Water-Demand Forecasting with Cross-Sector Historical Information. Water Resources Management, 40(14), 546. https://doi.org/10.1007/s11269-026-04912-9

Image Credits: AI Generated

DOI: 10.1007/s11269-026-04912-9

Keywords: water demand forecasting, XGBoost, machine learning, water–food–energy–ecology nexus, Taizhou China, regional water management, LSTM benchmark, Holt–Winters, walk-forward validation, demand-side analysis, Regional, Multi-Sector