Smart urban drainage systems could become a liability during communications failures
Decentralized control is resilient, but may be less optimal than centralized
Approximately centralized control enabled by distributed system estimates
Centralized control maintained in a case study when reporting every 7 days
Connectivity is automatically identified among water level sensors and valves solely using response data.
Inferred connectivity and response data are used to train linear feedback controllers.
Controllers built purely from data prevent flooding and greatly improve water treatment.
*First North American Winner of the Poul Harremoes Award for best urban drainage paper by a young author at the International Conference on Urban Drainage
Sensors spanning 200k km^2 built, deployed, and maintained by a small research group
Predictions automatically generated for sensors using only measurements and location
Efficient models train and predict for large networks using only a consumer laptop
Full architecture for predict-and-control sensor networks provided open-source
Process-based models are not designed to ingest large amounts of real-time data.
Machine learning approaches can automatically ingest data to make predictions, but lack interpretability.
A new open-source method automatically creates interpretable models, as validated with data from nearly 400 catchments.