Transforming Design Tools into Predictive Operational Tools through Artificial Intelligence and Machine Learning (AI/ML)

Our presentation will summarize the development and implementation of artificial intelligence and machine learning (AI/ML) tools Citizens Energy Group has developed to provide predictive operational decision support for combined sewer overflow (CSO) facilities. Citizens developed these tools by efficiently leveraging its prior investments in collection system modeling and metering. Our AI/ML tools consist of artificial neural networks fit to long-term datasets to predict inflows to CSO facilities using the 72-hour NOAA radar rainfall forecast. Both the rainfall and CSO facility inflow are embedded in a Power BI dashboard for efficient communication to Citizens staff. We developed this dashboard without incurring any additional software cost and believe any wastewater utility operating wet weather facilities can implement a similar AI/ML tool to ours. Citizens Energy Group is the combined water, wastewater, natural gas, steam, and chilled water utility in Indianapolis. Citizens is presently implementing a 20-year CSO consent decree with required completion in 2025. Consent decree compliance is primarily achieved through the DigIndy Tunnel system, a 28-mile long, 18-foot diameter tunnel network. Citizens is also implementing a pair of storage facilities in the Upper White River and Upper Pogues Run watersheds. Citizens, like all utilities implementing a consent decree, has made significant investments in modeling and metering its collection system, and has multiple years of meter data and model results on the shelf from recent regulatory and design support activity. While Citizens has confidence in its InfoWorks ICM model of the collection system, computational limitations prevent the model from being applied in real time or for immediate prediction of future rainfall events. In order to transform the collection system model from a design tool into a predictive operational tool, Citizens developed neural networks for the DigIndy Tunnel system and the Upper Pogues Run storage tank. These neural networks effectively serve as an extension of Citizens' collection system model to simulate inflow to the tunnel or tank in a matter of seconds. Since both facilities will begin operation at the beginning of 2022, Citizens utilized several years of collection system model input and output data to develop the neural networks. Figure 1 presents the architecture diagram for the AI/ML tools. Within Power BI, embedded Python scripts using the MetPy library (Unidata, 2021) connect to the NOAA server and obtain the 72-hour National Digital Forecast Database (NDFD) radar forecast for rainfall and temperature. The forecasted rainfall data is compiled for the CSO facility service area and temperature is utilized to estimate evapotranspiration (Hargreaves and Samani, 1982) for use in the neural networks. Citizens developed standard feed forward neural networks to predict inflow volume to the tunnel and storage tank similar to the collection system model. Figure 2 presents a scatterplot comparing each modeled inflow event to the tunnel system for precipitation years 2016 through 2019 to the neural network prediction. The inputs to the neural network are rainfall, peak intensity, antecedent dry days, the previous day's rainfall, and evapotranspiration. The total inflow volume from the neural network was within one percent of the InfoWorks collection system model simulation, with a R2 of 93%. Citizens evaluated input parameter sensitivity of the neural network and determined that for the storage tunnel, rainfall is the nearly an order of magnitude more important than any of the other input parameters. However, for the Upper Pogues Run storage tank that serves a much smaller watershed area, peak intensity is nearly as important as rainfall. Citizens initiated a desktop test deployment in early 2021 and progressed to an automated test deployment that has been completed at the time of the abstract. In the automated deployment, the rainfall forecast and subsequent inflow volume forecast is generated hourly, with results posted on Citizens' internal SharePoint. The rainfall forecast and forecasted inflow volume are presented though an online Power BI dashboard. We will present our method of deployment of this tool as well as other options for automated deployment that wastewater utilities may consider. Table 1 presents an example forecast from the neural network for July 15 – July 17, 2021. The forecasted inflow volume is based on the NDFD on July 14, 2021. In other words, should the tunnel system have been operational on July 16th, operational staff would have had an estimate of the magnitude of inflow more than a day before the event occurred. Over the course of the deployment, the BI dashboards were expanded to also collect NWS river stage forecasts and actual rainfall data through the application programming interface (API) provided by ADS as part of the PRISM interface to Citizens' rain gauge and flow meter network. Using the API key, Citizens developed a process to compare the forecasted rainfall from NOAA to the actual rainfall from the gauge network. From March 2021 through July 2021, the total forecasted rainfall was 23.5 inches, compared to 22.3 inches of actual gauge rainfall. An encouraging finding is that seventy-six percent (76%) of the rain events were forecasted within two-tenths of an inch of the actual gauge rainfall. We believe our presentation will benefit any wastewater utility that is operating or soon to be operating a wet-weather storage, treatment, or pumping facility to address combined sewer or sanitary sewer peak flows. We will provide a roadmap for any utility to implement a similar AI/ML solution for predictive operational decision support that can be implemented without any commercial software cost.