Development and Deployment of Data-driven Optimization Models in Full Scale Wastewater Treatment Plants

This presentation will describe a successful approach taken for developing and deploying 15 wide-scale machine learning algorithms in the wastewater and drinking water treatment industry. Attendees will learn the level of effort and investment required to truly leverage the power of operational data at scale and all the accompanying challenges, benefits, and considerations. The focus of the presentation will be on the technical implementation of Jacobs' Intelligent OM across eight full-scale treatment plants. Intelligent OM is a cyber-secure end-to-end cloud-based solution capable of integrating operational data and produce automated optimal dosing strategies to ultimately reduce excessive chemical usage or power consumption, while still meeting strict treatment and discharge requirements. The machine learning models discussed are developed by environmental engineers and data scientists and are unique to each treatment plant process train, operational limitations experienced by operators, and seasonal loads. The first model was deployed in April 2022 in Tucson, AZ with the goal to reduce aeration power required to meet nitrogen removal. The model was online continually until July 2024, going offline because of plant maintenance. Over that period the operators utilized the recommendations 65% of the time and yielded an estimated reduction between 15% to 20% in power usage.

Since then, 14 more unique models have been deployed across eight wastewater and drinking water treatment plants ranging from 10 to 200 MGD. These models target a combined 22 dosing parameters including ferric chloride in primary treatment, dewatering polymer, magnesium hydroxide for alkalinity control, sodium hypochlorite and sodium bisulfite for chemical disinfection, among others (Table 1). In addition to maintaining compliance and treatment goals, a suite of metrics are calculated to quantify the impact of the digital solutions, such as adoption rates, operator engagement and feedback, greenhouse gas emissions, and economic savings. Collectively 54% of all recommendations were implemented by the operators yielding an estimated $1.5M in savings and 574,000 kg CO2 equivalents diverted from chemical production emissions (Table 1). During the presentation the authors will discuss the technical challenges that are faced during production and deployment of these models, and where the focus is to maximize their impact.

The backend infrastructure is built upon Palantir Foundry, a data delivery software that enables cyber-secure connectors (Figure 1). All treatment plants in this work required custom connection protocols because of site specific SCADA, firewalls, software, and hardware security protocols. A data connector is configured with one-way communication and is limited to relevant information for developers. Additionally, it follows various safety practices such as encryption during transit or storage, user authentication and authorization controls. Once within Foundry, signals are processed in real time to assess sensor health and prepare data for the models. The architecture allows developers to have eyes on the system while giving operators and plant managers cyber-security peace of mind.

Accuracy of machine learning models is of upmost importance in digital solutions. For any given model, a training and test set are used to assess the accuracy of the regressions, which are generally built upon a handful of algorithms including random forests, multiple linear regressions, and gradient boosting trees. The end-to-end data pipeline (Figure 2) uses trained models to optimize control variables given current water quality conditions in the field. Taking dewatering polymer as an example use case, a neural network machine learning model was trained to predict the cake consistency given a handful of water quality parameters. This model achieved a root mean square error (RMSE) of 0.046 and 0.056 in the training and testing datasets respectively, meaning that the difference between predicted and actual values is low, therefore providing an accurate model (Figure 3). Finally, an analysis to measure the impact of polymer dose on cake solids percentage was run using a particle swarm optimizer. Using historical data as a training set, the optimizer was configured to minimize the polymer dose while maintaining 26% — 28% cake solids (Figure 4). In average a reduced dose of 2.6% was found. The model went live in May of 2023 and has yielded an estimated 5% reduction in polymer usage per dry ton.

The authors will cover the technical building blocks of digital solutions aimed for the optimization of any unit process. Emphasis will be given to the digital architecture and backend infrastructure that make full-scale deployments possible, the statistical backbone of machine learning models and optimization algorithms, and the results that can be expected once models are deployed and maintained.