{"resourceId":"reclamation-nwri-ai-ml-guidebook-potable-reuse-staged-validation-2026","versions":[{"version":"external-7e932d4bae7c3053a4149360ca31669fca15668c51f34b00cc05a950b980de0f","resource":{"id":"reclamation-nwri-ai-ml-guidebook-potable-reuse-staged-validation-2026","title":"Reclamation-funded AI & ML Guidebook for Potable Reuse sets staged, guardrailed validation for treatment-plant ML; newly promoted to U.S. utilities on 10 Nov","organization":"Carollo Engineers, Yokogawa, National Water Research Institute and Baylor University, for the U.S. Bureau of Reclamation","sector":"Drinking water, water reclamation and potable reuse treatment operations","geography":"United States (with a Japan-sponsored case study); applicable to conventional drinking water and wastewater plants","publishedAt":"May 2026 (final guidebook; day not stated, publicationDate set to first of month); publicised by ASCE Civil Engineering Source on 10 Nov 2026","publicationDate":"2026-05-01","eventDate":null,"sourceName":"National Water Research Institute (guidebook PDF)","sourceLabel":"Federally funded technical guidebook","sourceUrl":"https://www.nwri-usa.org/_files/ugd/632dc3_e4e9146bcb134de78c05d7c5e5020bf3.pdf","evidenceClass":"standards-guidance","outcomeClass":"mixed","topics":["data-security","infrastructure","operating-model","accessibility-workforce"],"finding":"The guidebook, from a Reclamation-funded project on data-driven fault detection and process control for potable reuse, gives utilities a staged path from desktop ML to full-scale use. It warns ML is 'extremely site-specific': models trained on one membrane or filter were inaccurate on others in the same plant, so tools proven at a demonstration facility need retraining at full scale. Safeguards: semi-autonomous trials (AI recommends, operators act) before automation, keep all prior alarms, trial on one parallel train, divert AI-controlled water to sewer for direct potable reuse, and program floors so AI cannot set treatment intensity below the old control system until months of validation. It flags data leakage (suspicious >99% accuracy) and calls for time-based splits. Case studies: a three-week full-scale dissolved-oxygen digital twin with 23% lower aeration energy per kg ammonia versus adjacent weeks; a modelled NDMA soft sensor where a two-standard-deviation safety factor halves UV energy savings from 26% to 13%.","sledRelevance":"Why now: ASCE promoted the guidebook to U.S. utilities on 10 Nov, quoting its author on phased implementation and data readiness. It is a rare U.S. federally funded, practitioner-written control framework for operational AI in treatment plants, directly usable in procurement and pilot design.","evidence":"Technical guidebook drawing on project work and cited studies. The 23% figure is a three-week, before/after comparison with adjacent weeks (no randomisation); the 26%/13% figures are modelled, not observed. ASCE's 'at least 10%' energy figure for a Southern California facility was not located in the extracted guidebook text and is attributed to ASCE only.","architectureImplications":"Requires robust data pipelines from SCADA/LIMS, model and data version management, and a control hierarchy where AI setpoints are bounded by the legacy control system.","governanceImplications":"Public-health protection plan before any AI control; reward developers on full-scale performance, not desktop accuracy; benchmark against a parallel train.","securityPrivacyImplications":"Cloud-hosted modelling (as in the digital-twin case) brings data transfer and OT-connectivity questions; the article notes utilities must consider cybersecurity, but the guidebook text inspected focuses on data and validation.","caveats":"Focus is potable reuse; transferability to conventional plants is asserted by authors. Engineering-firm authors with commercial interest; not peer-reviewed as a document. Older than the last run; included because it was newly promoted on 10 Nov.","streamIds":["utility-management"],"roles":{"sales":"Sales framing: treatment-plant managers are being offered ML optimisers with headline energy savings. This federally funded guidebook gives them a defensible way to test those claims: site-specific retraining, parallel-train benchmarks and floors that stop AI lowering treatment below current control. Position a bounded pilot-design or vendor-claim review using the guidebook's steps, not promised savings; its own case figures are short trials or models, and the safety factors needed cut savings substantially.","engineering":"Engineering framing: build data pipelines from SCADA and laboratory systems with version control, then validate models using time-based splits rather than random ones, and treat accuracy above 99% as a leakage warning. Deploy first as advisory recommendations, keep existing alarms, and test on one parallel train with setpoint floors tied to the legacy controller. Expect retraining when moving from pilot to full scale or between membranes. Review OT connectivity for any cloud modelling.","delivery":"Delivery framing: write a public-health protection plan before any AI touches setpoints, then run advisory trials, a single-train test, and staged expansion with sign-off at each step. Plant superintendent owns operational decisions; data science supports. Proposed acceptance criteria are measured energy or chemical savings against a concurrent control train, no compliance excursions, and documented performance after retraining. Risks are site-specific model drift, data gaps and staff capacity for maintenance."},"roleDetails":{"sales":{"customerProblem":"Treatment plants cannot easily judge vendor ML optimisation claims or deploy them without risking water quality.","stakeholders":["Plant superintendent","Water quality/compliance manager","Process engineer","State drinking-water regulator","Data/IT lead"],"discoveryQuestions":["Do you have parallel trains that could serve as a concurrent control?","How are SCADA and lab data stored and versioned?","What minimum treatment setpoints must never be breached?","How would you validate a vendor's model on your site?"],"valueHypothesis":"Applying the guidebook's staged validation could avoid unsafe deployments and produce site-specific evidence for or against an ML optimiser.","potentialEngagement":"Pilot-design and vendor-claim review using the guidebook's public-health plan, validation and parallel-train steps.","unsupportedClaims":["That a given plant will achieve 23% or 10% energy savings","That modelled 26%/13% UV savings were observed","That models transfer between plants or trains without retraining"]},"engineering":{"fit":"Drinking water, wastewater and reuse plants with SCADA historians and laboratory data, ideally with parallel trains.","architectureIntegration":"Advisory ML layer reading historian/LIMS data; any control setpoints bounded by the existing control system.","prerequisites":["Clean, time-stamped process and lab data","Parallel-train or comparable baseline","MLOps for model and data versions"],"constraints":["Models are site-specific","Regulatory approval for control changes in potable systems","Cloud connectivity limits for OT data"],"security":"Keep ML read-only from OT during advisory phase; review any cloud data transfer and access paths.","proofOfValue":["Time-based validation results","Concurrent parallel-train comparison","Performance after retraining at full scale"]},"delivery":{"work":"Create a public-health protection plan, data pipeline and staged trial from advisory to bounded control.","dependencies":["Regulator engagement","Data quality work","Operator availability"],"ownership":"Plant superintendent with process engineering.","skillsAdoption":"Train operators to interpret recommendations and data staff in model maintenance.","governanceCheckpoints":["Public-health plan approval","Advisory-to-control gate after months of validation","Post-retraining review"],"acceptanceCriteria":["Measured savings vs concurrent control train","Zero compliance excursions during trials","Documented setpoint floors enforced"],"risks":["Model drift and sensor fouling","Overstated desktop accuracy","Staff capacity to maintain models"]}},"retrievedAt":"2026-11-10T21:56:41.080910-06:00","enrichedAt":"2026-11-10T22:03:27.938369-06:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Cites an ageing workforce and the need for staff training and model-maintenance skills.","procurementImplications":"Pilot contracts can require time-based validation, parallel-train benchmarks and performance floors rather than desktop accuracy claims.","operatingModelImplications":"AI control authority expands only after months of full-scale validation; operators remain decision-makers during semi-autonomous trials.","peerReview":"unknown","sourceVerification":{"openedUrl":"https://www.nwri-usa.org/_files/ugd/632dc3_e4e9146bcb134de78c05d7c5e5020bf3.pdf","referenceExcerpt":"Program constraints such that the AI/ML cannot set the treatment intensity lower than the old control system, or another safe minimum","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}