A new guidebook is offering water utilities a practical roadmap for introducing artificial intelligence and machine learning into potable reuse operations, while warning that successful adoption will depend as much on data quality, cybersecurity and workforce readiness as on the technology itself.
The AI & Machine Learning Guidebook for Potable Reuse was developed through a U.S. Bureau of Reclamation-funded research project led by Carollo Engineers in collaboration with Yokogawa, the National Water Research Institute (NWRI) and Baylor University.
The free guidebook is intended to help utilities move from identifying potential artificial intelligence (AI) and machine learning (ML) applications through model development, pilot testing, full-scale deployment and long-term maintenance. Although focused on potable reuse, its authors say many of the principles can also be applied to conventional drinking water and wastewater treatment.
The guidebook comes as utilities generate growing volumes of operational data through supervisory control and data acquisition (SCADA), metering and Internet of Things sensors, but often lack the staff, systems or expertise to turn that information into operational insights. The report notes that meaningful ML adoption can require investments in sensors, data management, model development and ongoing retraining, while the benefits can be highly site-specific.
Subscribe to our Newsletter!
The latest environmental engineering news direct to your inbox. You can unsubscribe at any time.
“Artificial intelligence and machine learning have the potential to help utilities make better use of the vast amount of operational data generated every day,” said Andy Salveson, vice president at Carollo Engineers and principal investigator for the project, in a statement. “This guidebook provides a practical path forward for utilities interested in exploring these technologies, from understanding the fundamentals to successfully implementing and maintaining tools that support more informed operational decisions.”
Guide highlights real-world water applications
Potential applications identified in the guidebook include multivariate water quality alerts, “soft sensors” that use available data to estimate parameters that may not be measured continuously, water quality and flow forecasting, predictive maintenance and digital twins.
Several case studies demonstrate how those approaches have already been tested.
In one Water Research Foundation project, supervised ML was used to identify industrial contamination spikes entering a potable reuse system. Compared with conventional fixed-threshold alarms, the ML approach detected events sooner and generated fewer false positives.
Another case study describes a full-scale digital twin developed by Carollo and Yokogawa for a water reclamation facility upstream of a potable reuse demonstration facility. The system calculated an ideal dissolved oxygen setpoint and provided hourly recommendations to operators. During a three-week trial, aeration energy use per kilogram of influent ammonia was 23% lower than during adjacent weeks.
The guidebook also describes an ML model used to forecast flows from a membrane bioreactor feeding three reverse osmosis trains. Predictions remained accurate for up to 48 hours and were used to develop a proactive schedule for starting and stopping RO trains, helping the utility work toward an annual reclamation target of approximately 492,100 cubic metres (130 million gallons).
Other research highlighted in the document includes the use of unsupervised ML to analyze California PFAS data. Researchers found distinct patterns in biosolids data that could potentially help identify PFAS sources, including a cluster associated with wastewater treatment plants located closer to landfills.
Data management and cybersecurity are key hurdles
The guidebook emphasizes that utilities should not begin with AI technology itself, but with a clearly defined operational problem and an assessment of whether ML is the appropriate tool. Its implementation framework begins with problem definition before moving through data collection, model development and eventual scale-up.
Data management is identified as one of the most important foundations for successful deployment. The report recommends breaking down data silos between operations, laboratory and maintenance departments and connecting sources such as SCADA and laboratory information management systems into unified data pipelines. Models should then be validated against new field data and monitored for declining performance or “model drift.”
Cybersecurity also needs to be incorporated from the beginning of an AI or ML project, according to the guidebook. Connecting models with sensors, SCADA, programmable logic controllers and operational technology can create new points of vulnerability during data collection, model training, deployment, retraining and ongoing operation.
The report recommends a phased approach in which models are initially deployed in controlled settings, such as a pilot or parallel treatment train. Utilities should quantify benefits such as reductions in energy use, water loss, operator hours or chemical consumption before expanding a system. Existing control systems should also remain available as backups as AI or ML applications move toward full-scale use.
Long-term staffing may be bigger challenge
Keeping an ML system operating over decades could prove more difficult than developing the initial model, the report cautions. Seasonal changes, equipment upgrades, staff turnover and changing operational priorities can all affect performance. Long-term success therefore depends on organizational capacity, workforce readiness and strategic planning in addition to model accuracy.
Staffing presents a particular challenge. The guidebook notes that specialized roles such as data scientists and ML operations engineers remain uncommon at water utilities, meaning many systems may need to rely on consultants, technology vendors, universities or regional partnerships for more sophisticated applications.
Operator participation will also be critical. The guidebook recommends involving operators early in system design, establishing formal feedback mechanisms and ensuring that AI and ML tools supplement rather than simply replace human judgment.
The AI & Machine Learning Guidebook for Potable Reuse is available free to utilities, regulators, researchers and other water professionals.
