Accurate water demand forecasting is essential for the operation of drinking water distribution systems. Although machine learning approaches have demonstrated superior predictive performance in research settings, their adoption by water utilities remains limited. A primary barrier is the translation from model development to productive deployment, which includes model deployment, monitoring, retraining, and infrastructure provisioning. This paper presents an end-to-end MLOps architecture for operational short-term water demand forecasting that addresses these deployment challenges. The proposed system integrates established open-source components into a scalable Kubernetes-based framework supporting the complete model lifecycle as well as the ML pipeline reaching from data ingestion through inference to forecasting. As exemplary approaches, DeepAR models with a minimal feature set derived from flow measurements, meteorological observations, and temporal encodings are used. The application of the architecture is demonstrated on two distinct water distribution systems: a municipal network in southern Germany requiring 24-hour forecasts, and a district metered area network in northern Italy with weekly prediction horizons. The analysis reveals that the test performance provides only a partial indication for operational ML pipeline forecast quality. Key findings include the necessity of robust gap-handling strategies for event-driven data streams, the importance of continuous performance monitoring and adaptive retraining, and the value of probabilistic forecasts for uncertainty-aware decision support.
Wünsch et al. (Thu,) studied this question.