Robust Machine Learning Pipeline for Water Usage Prediction and Systematic Water Management

• Activity Number: 561 - Statistical Analyses for Environmental Monitoring
• Type: Contributed
• Date/Time: Thursday, August 11, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Jimmy Yu and Jeffrey Blake Bullwinkel and Jason Parcon and Lochana Kanishka Palayangoda and Anthony Weishampel*
• Companies: PepsiCo Inc and Harvard University and PepsiCo Inc and PepsiCo Inc and PepsiCo Inc
• Keywords: Water Usage Ratio; Water Conservation; Time Series; Anomaly Detection; Clutering; Forecasting

Abstract

The proliferation of smart water meters that measure water consumption has enabled organizations to collect water usage time series for a wide range of water-intensive processes, including agriculture and industrial manufacturing. In many cases, however, these organizations lack the ability to understand their water usage at a given point in time, set reasonable goals for future consumption, and identify major water-losing events before they happen. We present a series of methods for anomaly detection, clustering, and forecasting to address these tasks using an automated pipeline, providing insights that enable data-driven water management and warnings that mitigate water loss. Our pipeline offers deeper insights with more granular data but is flexible to a range of temporal scales. We demonstrate its efficacy using proprietary weekly water consumption data collected by PepsiCo Beverage North America (PBNA) plants and achieve 95% average accuracy on anomaly detection and 43% average reduction in the RMSE of next-week forecasts in comparison to a moving average baseline.