Enabling Uncertainty-Aware Time-Series Forecasting in Federated Learning for Urban Water Dynamics

Urban water forecasting is essential for flood awareness and resilient city operations. Yet, in real-world deployment, sensor data is fragmented due to utility privacy policies and sensor failures. Therefore, city-scale forecasting must operate across fragmented sensor ownership while providing calibrated uncertainty under sensor failures. To address this, we extend the functionality of the transformer forecasting model by introducing i) a quantile regression head for uncertainty calibration on forecasting intervals, and ii) a sensor-disjoint federated training scheme, scaled up to 1,000 sensors across 100 clients without raw-data exchange in between, with remaining robustness to a Zipf-1 ownership skew. The quantile head preserves point-forecast absolute error from 0.0489 to 0.0338. The federated training retains this skill end-to-end. The developed framework is easily adjustable to local privacy policies, provides uncertainty-aware forecasting at its backbone, and applies to city-scale risk management. We map the architecture developed onto the canonical world-model recipe and identify the missing components for closed-loop urban water management.