Machine Learning and SMART Sewer Management

Various machine learning techniques have been applied in collection system model calibration, optimization, alternative analysis and operational controls in recent years. Various tools have been developed for these applications accordingly. The growing experiences gained with these applications made it possible to combine these tools into an integrated platform for comprehensive SMART sewer management.

Jacobs initially developed AutoCal for the purpose of calibrating large hydrologic and hydraulic models. When applied to a model, the tool predicts the change in calibration parameters necessary to produce the best representation of the observed data with the model results – efficiently calibrating the model with an automated workflow.

Because AutoCal learns model input and output relationships, it can be flexibly applied to a variety of engineering problems and models, provided the necessary modules and connections are built out in the tool. Several AutoCal variations have been developed to perform tasks such as system operation optimization and cost effective or best performance alternative analysis. Besides the SWMM engine, AutoCal is also configured to interface with a number of hydrologic and hydraulic models such as RORB, TUFLOW, CHI PCSWMM, Mike Urban, and InfoWorks ICM. It has also been applied to the groundwater model MODFLOW demonstrating the wide application of surrogate optimizers such as AutoCal.

In 2020, Jacobs began operating the City of Wilmington’s wastewater treatment system, which includes the wastewater treatment plant (WWTP), the CSO control facilities, and the wastewater pump stations. Previously, only 6 out of the 41 CSOs had level and/or flow monitoring. The CSO monitoring expansion project was developed to add level monitors at the 35 unmonitored CSOs as well as flow meters on three interceptors and integrate them with the existing monitoring system into a state-of-the-art intelligent operations platform.

The collected data were integrated along with rain gauges, RTC facility and pump station SCADA data, tide gauges, tide forecasts, and weather forecasts onto one intelligent operations platform called Aqua DNA. With Aqua DNA, machine learning was applied as data were collected in real-time to learn the behavior of the system, developing a “baseline” level at the CSOs. Alerts were then developed for “off baseline” as well as if data transmission stopped or if overflows were occurring. The “eyes on the system” dashboard provides a view of the severity of the alerts on a GIS map that allows the CSO crew to know where their biggest issues are occurring in real-time. The platform also uses machine learning and artificial intelligence to predict the behavior of the system, which enables operators to see when a dry- or wet-weather CSO event is imminent and use this new insight to make operational decisions at pump stations and the WWTP.

The AutoCal tool is incorporated into the platform. The real-time data are being used with the autocalibration module to provide near real-time calibration/refinement of the hydraulic model. Additionally, the optimization module is set up to run rainfall forecasts, especially for smaller and medium-level storms, to optimize setpoints at the RTCs and pump stations to maximize flow to treatment and minimize upstream SSOs and downstream CSOs.

The solution has integrated the hydraulic model to estimate flows from the City with a data-driven model to estimate flows from the County for which the City does not have a hydraulic model. By combining these estimates, decisions can be made to optimize operations ahead of the storm, including making sure that the City’s terminal pump station is pumping at full capacity for as long as possible. Additionally, running the model allows for the operators to see how each CSO will react to the impending storm, even showing where levels might increase after the storm is over.

In conclusion, the machine learning technique, i.e., “eyes on the system”, has informed the City and the CSO crew of potential problem areas that were not previously known and helped to reduce the crew’s drive time and traffic control. The CSO crew are now able to be proactive in their maintenance, rather than reactive. The combination of the hydraulic model and data-driven models have allowed the WWTP operators to know what flow may be headed their way ahead of the storm. Decisions can now be made to maximize treatment and minimize SSOs and CSOs by reviewing the results holistically.

Learning Objective:

This presentation presents applications of machine learning techniques in collection system modeling, and a case study for smart combined sewer overflow monitoring and control project in Wilmington, DE.