Machine-Learning-Water-Systems-Model

This machine learning workflow demonstrates a framework to function a digital twin of a systems dynamics model for urban water system seasonal water system reliability, resilience, and vulnerability analysis.

Abstract

Reliability, Resilience, and vulnerability (RRV) of water systems inform operational and management decisions regarding the system's supply and demand interactions. Systems dynamics modeling can replicate the physical processes to evaluate system performance, but their high parameterization, high computational requirements, immense development period, difficulty calibrating, and software licensing present significant barriers to their widespread integration in practice and inhibit translation of research to practice. To address these challenges, this study utilizes the Xtreme Gradient Boost (XGBoost) algorithm, automated feature selection, and automated hyper-parameter optimization as a novel machine learning framework to predict daily reservoir levels, groundwater extraction rates, and out-of-district water requests for determining each their RRV. We examine the XGBoost water systems model (XGB-WSM) forecasting accuracy during dry, average, and wet climate scenarios compared with a water systems model developed to assess Salt Lake City's water system vulnerability to climate and population growth. The XGB-WSM accurately models seasonal reservoir level dynamics, groundwater extraction rates, and out-of-district requests during the dry and average climate scenarios with a low RMSE and high R² (>0.91). The model accurately predicted all water system components for all hydroclimate scnearios, correctly demonstrating the seasonal trends and relationships to water system component thresholds. We find that machine learning demonstrates high potential for further development and integration in water resources planning and management, such as identifying and optimizing system operations, increasing community engagement, and strengthening the understanding of the water system for utilities without an existing systems model.

About the Salt Lake City Machine Learning Water System Model

This ML-WSM leverages decades of collaborate water systems modeling development between the University of Utah, the University of Alabama, and the Salt Lake Department of Public Utilities (SLCDPU), leveraging an existing documented and calibrated systems dynamics model (Salt Lake City Water Systems Model, SLC-WSM) built in the Goldsim software package for machine learning model training and testing. This model serves as a template for other water systems with or without an existing systems dynamic water systems model to adapt and create a machine learning water systems model tailored to their unique system interactions and feedbacks.

Brief Salt Lake City water system background

Study Area: Salt Lake City, Utah depends on winter snowpack in the adjacent Wasatch mountains to support surface water supplies, fill the Dell reservoir storage system, and replenish valley aquifers. Its mountainous topography, geographical location, and arid climate result in highly skewed April to October water use to counteract seasonally high evapotranspiration for outdoor landscaping and irrigation.

The SLCDPU shares many similarities with western and intermountain water utilities in growing metropolitan areas. The region's interannual climate variability and seasonality strongly influence winter snowpack accumulation, extent, and duration, functioning as the primary mechanism controlling surface water supplies. The region's cold semi-arid (BSk) to cold desert climate (BWk) significantly influences seasonal water use in response to high evapotranspiration during summer months. From April to October, outdoor water use approaches 1000 mm to irrigate commercial and residential landscaping. While a supply limited region, high seasonal water use places Utah as the 2^nd or 3^rd highest per-capita water use state depending on the year.

The utility satisfies its demands by sourcing adjacent Wasatch mountain surface water, underlying valley groundwater, and nearby Deer Creek reservoir and Central Utah Project (CUP) out-of-district supplies. From the Wasatch mountains, City creek, Parley's creek, Big Cottonwood creek, and Little Cottonwood creek contribute over 60% of the municipality's annual supply. The Parley's watershed contains Mountain Dell reservoir and Little Dell reservoir that hold up to 3.2x10⁶ m³ and 25x10⁶ m³, respectively, and are the only utility owned long-term storage sources. When surface water supplies cannot meet demand during the summer months, SLCDPU has access up to 22x10⁶ m^3</sup> per year of groundwater. If surface and groundwater supplies cannot satisfy demands, out-of-district Deer Creek reservoir and CUP sources support up to 61x10⁶ m³ per year of use. This water comes at a greater cost, resulting in it being the least prioritized supply source, and for purposes of RRV metrics in this study, considered a reliability failure for the water system.

Water System: Salt Lake City water system leverages adjacent Wasatch Mountain surface water supplies, small reservoirs (Dell system), groundwater withdrawal, minimal in-system storage, and access to larger U.S. Bureau of Reclamation reservoir systems.

Machine Learning Model Inputs and Training

Many of the SLC-WSM inputs drive the XGB-WSM, including daily surface water supplies (e.g., City Creek), total system demand, service area population, reservoir levels, and the previous time step's reservoir levels, groundwater extraction rate, and out-of-district requests. For example, the reservoir level on July 1^st functions as an input to predict the reservoir level on July 2^nd. This research developed three additional metrics to further enhance model performance: total daily surface water supplies, day of the year, and month. For model training we source streamflow and demand data from the utility's long-term records, the Kem C. Gardner Policy Institute provides population data, SLC-WSM simulations provide groundwater extraction rates, out-of-district Deer Creek reservoir use, and Mountain Dell and Little Dell reservoir levels. Streamflow values entering water treatment facilities at each canyon's mouth form model inputs, as well as demand data conisting of the total volume of water entering the distribution system (all connected demands, leaks, and unaccounted-for losses).

For model demand predictions, the ML-based Climate-Supply-Development water demand model (CSD-WDM, https://github.com/whitelightning450/Water-Demand-Forecasting) predicts mean monthly demands in response to climate, supply, and socioeconomic factors. The CSD-WDM demonstrates high seasonal prediction accuracy with a mean absolute error of 62.8 lpcd and a mean absolute percent error 8.4%. Since XGB-WSM operates at a daily time step, cubic spline interpolation downscales the demand predictions to a daily temporal resolution of total system water demand.

The XGB-WSM uses these features to predict daily Mountain and Little Dell reservoir level, system groundwater extraction rate, and out-of-district Deer Creek reservoir requests during peak SLCDPU water use between April and October. Model training is on seventeen years of daily simulation data spanning from 2000 to 2020, omitting the three testing scenarios described below.

Evaluation Scenarios

XGB-WSM predictive performance is on three different scenarios, based on annual snowpack, to represent the intermountain regions' hydroclimate variability. The Alta Guard MesoWest weather station in the headwaters of Little Cottonwood Creek provides a long-term (1945-present) snowfall record to identify the most recent dry (2015), average (2017), and wet (2008) conditions for XGB-WSM simulation evaluation. A Log-Pearson Type III analysis indicates the dry year demonstrates a non-ex