Holistic Wet Weather Management Combining Machine Learning, Treatment Plant Optimization, and Predicting Collection System Influent Flow Hydrographs

Purpose

The purpose of this paper is to describe a machine learning (ML) approach that was implemented to optimize wet weather operations at the Neuse River Resource Recovery Facility (NRRRF). This was a collaborate effort between Hazen and Raleigh Water to investigate if a ML approach could be used to predict influent flow to the NRRRF 72-hours in advance as a function of explanatory variables with an emphasis on being able to accurately forecast the largest events. Benefits The project objectives were achieved, and this presentation will demonstrate that many facilities can follow this approach to predict the shape of their influent flow hydrograph and implement their wet weather strategy accordingly in real-time. This paper will also include recommendations for future projects such as including ongoing model retraining to more accurately and quickly incorporate changes in collection system behavior into your model.

Introduction
Raleigh Water owns and operates the NRRRF, which treats an average daily flow of 48 million gallons per day (mgd), is permitted to treat 75 mgd, and can convey a hydraulic peak hour flow of 225 mgd. In the past few years the NRRRF has experienced extended 24-hour sustained flows around 150 mgd and peak hour flows around 184 mgd during wet weather. The NRRRF is subject to a stringent total nitrogen (TN) load allocation established under the Neuse River Basin Nutrient Management Strategy that requires the Neuse River RRF to achieve an annual average TN concentration of less than 3 mg/L at permitted flow. In addition, NRRRF must also meet a quarterly average effluent total phosphorus limit of 2 mg/L. The biological nutrient removal process's performance is typically inversely correlated to the amount of flow and load entering the NRRRF, hence the ability to minimize the peak flow through the facility would improve effluent quality and process reliability. The NRRRF's 32-million-gallon primary effluent equalization basin was designed to withhold a significant portion of the flow and load entering the NRRRF during high flow events. While the equalization volume is substantial and offers a useful tool for operators, its utility could be optimized by knowing how much flow will enter the NRRRF for a given rainfall event over the duration of the event and how high the peak flow will be. However, even if a facility does not have dedicated equalization basins such a predictive tool can still be beneficial as other wet weather strategies can be planned for and implemented. Prior to this project, NRRRF Staff utilized collection system pump station data to estimate when the peak flow of the storm will reach the NRRRF and its magnitude, which provided about 30-60 minutes of advance warning. Raleigh Water also has a collection system model and collection system flow monitors. The collection system model is useful planning tool, but there is currently not a way to use this tool in a real-time fashion. ML uses algorithms that assign weights to independent variables and seeks to minimize error when predicting a dependent variable. The open source software package Python was used in this evaluation. ML is an alternative to traditional mechanistic models. The project outcome described herein (predicting influent flow 72 hours in advance) could likely be achieved using a well-calibrated collection system model and forecasted rainfall data as well. Some of the advantages of the ML approach are that it utilizes non-proprietary software, can be deployed to provide continuous retraining, and the results can be viewed in customized format that integrates with existing tools such as the NRRRF's secondary clarifier guidance program. This is not to say that ML models are superior or should replace traditional scientific approaches.

Results
A well-calibrated model predicting influent flows to the NRRRF was developed using a ML-based approach. A ML model was trained to over six years of hourly influent flow data to predict influent flow using the following explanatory variables: the past 12 hours of influent flow, streamflow data, and rainfall data. The model utilizes hourly rainfall forecasts and real-time streamflow data in its predictive algorithm. It is worth noting that collection system monitoring data was also considered but was unavailable for the same time period as the rest of the data, but this information ought to be considered if available. Models were developed to predict flow at each time step from 1 to 72 hours into the future. Thirty-eight storms in 6 years of training data were accurately predicted. The predictions are displayed in a web-based Microsoft Power BI dashboard tool that includes a tool to estimate the optimal point to fill the equalization basin to maximize its utility (see Figures 1 and 2). Microsoft Azure was used to develop an automated data pipeline to update model predictions each hour. The project was deployed in a test mode in December 2019 and completed in July 2020. Since then, at least two major storm events including Hurricane Isaias have occurred and been well predicted (see Figures 3 and 4). This tool also integrates with the secondary clarifier guidance program that calculates required clarifier surface area as a function of SVI, influent flow, RAS flow, and mixed liquor concentration. Per the original design the equalization basins were provided to limit the maximum day flow to the secondary system to two times the influent design flow, or 150 mgd. Use of the equalization tank during high flow events is important for maintaining sustained flows to the secondary process below the capacity of the clarifiers. This is also why the 24-hour running average flow is displayed on the model prediction screen.

Conclusion
Machine learning tools have a place in the water industry and are likely to become more commonplace in the next decade.