Abstract: Comprehensive seasonal water system assessments in the western U.S. are confounded by per-capita water demand stationarity, owing to increased system uncertainty from inter-annual demand variability not represented in industry forecasting methods. Using Salt Lake City's Department of Public Utilities observed demand variability from recent episodes of drought, average, and surplus supply conditions, we measure the prediction accuracy of industry demand forecasting methods embedded with per-capita stationarity assumptions. To address the observed variability in intra- and inter-annual water demands, an ensemble of machine learning variable selection and optimization tools, hydro-climate and exogenous service area factors, and regression algorithms are leveraged to form the Climate-Supply-Development Water Demand Model (CSD-WDM). A multi-layered perceptron (MLP) and Random Forest regression (RFR) model are also developed to evaluate model complexity vs. accuracy trade-offs with the CSD-WDM. Industry forecasting methods exhibited high prediction errors in all climate scenarios, which we attribute to insufficient model complexity. These errors peak during drought conditions where over-predictions near 90% and 40% of the observed monthly and seasonal demands, respectively. The mean absolute percent error of CSD-WDM (8.4%) shows improvement over the MLP (9.0%), RFR (10.5%), and industry models (31.0%), accurately projecting demands in all climate scenarios. Overall, the CSD-WDM reduced seasonal water demand uncertainty by 6.0%, 40%, and 30% during surplus, drought, and average climate conditions when compared with industry methods.
This repository contains the data (in the Data branch) and modeling scripts (also containing data analysis and figure development) for the Salt Lake City Water Demand Model (SLC-WDM), multi-layered perceptron (MLP), and Random Forest (RFR). Each of these scripts also contains monthly stationary methods for comparision of forecasting performance and analysis.