Development and Performance of a ML-Based Tool That Predicts Influent Flow to a WRF 72-Hours In Advance And Integrates With Existing Wet Weather Nutrient Management Protocols.

Abstract

Influent flow to a 75-mgd Neuse River Resource Recovery Facility (NRRRF) was modeled using machine learming (ML). The trained model can predict hourly flow 72-hours in advance. This model was deployed in July 2020, and has been in operation over two and a half years. The model's mean absolute error (MAE) in training was 2.6 mgd, and MAE has ranged from 10-13 mgd in deployment for any point during the wet weather event when predicting 12-hours in advance. As a result of this tool, plant staff have optimized use of their 32 MG wet weather equalization basin, using it approximately 10 times and never exceeding its volume.