AECOM pitches open-source predictive model to county procurement committee for pipeline failures

AECOM representatives presented to the county procurement committee overseeing solicitation RPS 25-00208 in a virtual oral presentation that the firm can deliver a platform-agnostic, open-source predictive model for pipeline failures and a user interface the county would own and operate.

Why it matters: The committee was told the model would allow the county to shift from reactive repairs to data-driven capital planning and condition-assessment prioritization, and that the approach avoids subscription “black boxes” by keeping algorithms open and auditable.

AECOM’s presenters — lead Sameer Alkati, technical lead Chris Macy and project manager Tania Simmons — described a multistep program that reuses the county’s existing Esri, Oracle and Microsoft investments, uses Monte Carlo scenario analysis and produces outputs intended to drive rehabilitation and replacement prioritization. Alkati said the proposal favors “no black boxes” and that the firm’s algorithms are “fully open source. They’re explainable and auditable,” enabling local staff to defend and adapt the models to future needs.

Chris Macy outlined the technical rationale for the modeling approach, saying modern predictive models should model the deterioration and intervention processes and that different pipe sizes and materials require different treatments. Macy described three broad model types used in the industry and warned that some machine-learning techniques — for example, neural networks — can fit historical data well but perform poorly in forward-looking prediction. He recommended hazard-function approaches (for example, Weibull fits) and cohort screening by pipe era, pressure and material to produce explainable results. Macy cited examples from the City of Toronto and Colorado Springs where AECOM models guided capital planning and condition-assessment programs.

Presenters gave concrete assumptions and delivery timing. AECOM said an accelerated schedule of about 6–8 months is feasible by overlapping tasks; it said 2–3 months per data review/cleaning step is assumed “if the quality of your data is relatively good,” a condition the team attributed to the county’s mature asset-management program. The firm also said it has modeled roughly 60,000 miles of pipe in prior engagements and uses scenario analysis to show how different investment strategies change future failure counts.

Committee members asked technical questions during a 15-minute Q&A. Questions and AECOM’s responses focused on three areas: the handling of decadal spikes in installation/failure records, how the county should measure model success, and whether the delivered tool would remain dynamic and updatable.

On decadal spikes — records in the county data that cluster by decade rather than by year — Macy said that, for small-diameter pipes, the failure data themselves can inform era-distribution and that models can be stabilized to proportion failures reasonably over the decade. He said unknown or lumped dates are less problematic when failures are numerous, but are more consequential for materials whose deterioration is highly era-dependent, such as ductile iron and asbestos-cement pipe. For thermoplastics, Macy suggested opportunistic sampling and laboratory testing to determine extrusion quality when that material is suspected to be a significant risk.

On success metrics, Macy emphasized forward-looking validation: he recommended periodic recalibration and monitoring against future failures rather than measuring accuracy only against historical fits. He warned that some accuracy metrics (for example, predicting pipes that do not fail) can be misleading because most pipes do not fail in a given period; instead the model should be judged on its ability to predict and prioritize likely failures and to improve the effectiveness of condition-assessment programs.

On deliverables, AECOM said the county would “own it, you’re gonna own it, the model,” and that the deliverable is portable and built on existing systems rather than a separate commercial platform. Tania Simmons said the firm would work with county staff to configure the dashboard and user interface — for example, in Power BI — consistent with the county’s current tools and reporting preferences.

No formal decision or procurement award was made at the meeting; the session was an oral presentation for the RPS 25-00208 solicitation and concluded with instructions that the recorded presentation will be posted online. The committee asked clarifying questions but did not take votes or issue directions recorded as binding actions.

The presentation included multiple references to prior AECOM work in jurisdictions including Toronto, Colorado Springs, DeKalb County and Tampa Bay Water; presenters said those engagements support the firm’s claims about model portability and long-term performance. Presenters also reiterated that model performance depends on data quality and on periodic calibration following deployment.

For follow-up: the committee will review the posted recording and the procurement process will continue per the RPS 25-00208 schedule (no committee direction to staff or contract award was recorded during the presentation).